Spinal disease image segmentation technology integrating U-ResNet and shape-aware attention.

Spinal disease image segmentation technology integrating U-ResNet and shape-aware attention - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 13;16:12465. doi: 10.1038/s41598-026-42870-9 Search in PMC Search in PubMed View in NLM Catalog Add to search Spinal disease image segmentation technology integrating U-ResNet and shape-aware attention Dexuan Zhao Dexuan Zhao 1 The Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, 222000 China Find articles by Dexuan Zhao 1 , Rujie Qin Rujie Qin 1 The Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, 222000 China Find articles by Rujie Qin 1, ✉ , Zhijin Chai Zhijin Chai 2 Lianyungang Clinical College of Nanjing Medical University, Lianyungang, 222000 China Find articles by Zhijin Chai 2 , Shaoxin Ma Shaoxin Ma 1 The Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, 222000 China Find articles by Shaoxin Ma 1 , Qipeng Gao Qipeng Gao 1 The Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, 222000 China Find articles by Qipeng Gao 1 Author information Article notes Copyright and License information 1 The Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, 222000 China 2 Lianyungang Clinical College of Nanjing Medical University, Lianyungang, 222000 China ✉ Corresponding author. Received 2025 Dec 16; Accepted 2026 Feb 27; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13083956  PMID: 41820475 Abstract The incidence of spinal diseases is rising and affecting younger people, making early and accurate diagnosis based on medical imaging crucial for treatment. However, traditional manual segmentation and measurement suffer from issues such as judgment discrepancies, time-consuming processes, and subjective errors. Existing deep learning segmentation methods still face challenges such as insufficient adaptability to complex pathological interference, high deployment barriers, and weak integration with clinical needs. This study designs an end-to-end deep learning model, with three core optimization modules: a customized U-ResNet backbone network that balances feature extraction depth and computational efficiency through multi-scale feature fusion strategies to adapt to structural differences in different spinal segments; a shape-aware attention module that integrates semantic features and contour prior features to enhance the ability to capture changes in spinal morphology and suppress background interference; and a dynamically weighted combined loss function that adjusts the weights of region and boundary losses based on vertebral body and intervertebral disc characteristics, integrating clinical constraints to meet quantitative diagnostic needs. Experiments were conducted on the Lumbar Spine MRI and VerSe datasets, and the results show that the model outperforms existing mainstream models in classification, segmentation, and intervertebral disc degeneration grading tasks. This study can provide a reference for the design of task-oriented models for medical image segmentation, and can also provide technical support for the hierarchical diagnosis and treatment of spinal diseases in primary hospitals, helping to improve diagnostic efficiency and accuracy and reduce the medical burden. Keywords: Spinal disease segmentation, U-ResNet, Shape-aware attention, Medical image processing, Dynamic weighted loss function Subject terms: Computational biology and bioinformatics, Engineering, Health care, Mathematics and computing, Medical research Introduction The spine serves as the core structure for human movement and bodily support, with a direct impact on the physiological functions of the entire body. The incidence of spinal diseases–including lumbar disc herniation and cervical spondylotic myelopathy–has been on a continuous upward trend. These diseases not only exhibit high prevalence among middle-aged and elderly populations but also have spread to office workers and students who maintain prolonged sitting postures, presenting a clear trend of younger onset 1 , 2 . Early and accurate assessment of such spinal diseases is crucial for preventing severe complications, and it plays a significant role in improving the overall outcomes of clinical treatment. Spinal image segmentation faces unique challenges that hinder clinical application. The spine consists of multiple complex structures, and pathological changes such as fractures and deformities further increase segmentation difficulty. Traditional machine learning methods including threshold segmentation and region growing rely on manually designed features such as grayscale values and textures, struggling to address uneven grayscale and partial volume effects in medical images 3 – 6 . Existing deep learning methods also have limitations: Transformer-CNN hybrid models suffer from excessive parameters, and common loss functions like Dice loss fail to handle boundary differences between vertebrae and discs 7 , 8 . The U-Net network marked the deep learning era in medical image segmentation, addressing traditional method shortcomings via its encoder-decoder structure and skip connections that fuse shallow details and deep semantic features. It has become the fundamental architecture for spinal segmentation, with subsequent iterations improving accuracy through feature extraction, region focusing, and loss optimization 9 – 11 . However, U-Net lacks targeted optimization for spinal tasks: its fixed feature extraction cannot adapt to multi-segment spinal clarity differences, it fails to capture pathological morphological changes, and its conventional loss leads to poor vertebrae-disc boundary accuracy. To tackle these U-Net limitations in spinal segmentation, we first optimize the network’s core architecture while preserving its feature fusion advantage. We replace U-Net’s original backbone with a customized U-ResNet, integrating multi-scale feature fusion and dynamic receptive field convolution. This modification balances feature extraction depth and computational efficiency, allowing the network to adapt to structural differences across cervical, thoracic, and lumbar segments. We also enhance U-Net’s skip connection mechanism by adding modal adaptive normalization, which standardizes feature distributions from MRI and CT images–resolving U-Net’s poor adaptability to multi-modal spinal data. Building on this optimized U-Net foundation, we further develop a complete end-to-end model tailored to clinical spinal segmentation needs. We introduce a shape-aware attention module that fuses semantic features from the optimized U-ResNet decoder with contour prior features from preprocessing, enhancing the model’s ability to capture pathological morphological changes. We also design a dynamically weighted combined loss function that adjusts the weights of region loss and boundary loss based on vertebral and disc characteristics, addressing U-Net’s residual boundary segmentation defects and meeting clinical quantitative diagnosis requirements. This paper makes the following contributions: We propose a customized U-ResNet backbone (for U-Net optimization) designed to address the clarity differences between various spinal segments, improving both segmentation accuracy and computational efficiency. We introduce a shape-aware attention module that enhances segmentation accuracy by incorporating shape-related features and addressing pathological variations in spinal morphology, complementing the optimized U-Net. We design a dynamic weighted combination loss function that adjusts the weight between Dice loss and boundary loss based on the unique characteristics of vertebral bodies and intervertebral discs, optimizing both global segmentation and local boundary accuracy for spinal tasks. This thesis has six chapters. Introduction: research background clinical needs main innovations. Related Work: existing spinal image segmentation research technology challenges advancements. Methods: model structure U-ResNet backbone shape-aware attention dynamic combined loss. Experiments: datasets evaluation metrics results model effectiveness. Results and Discussion: performance analysis baseline comparisons ablation studies. Conclusion: findings summary future work directions. Related work Diversified optimization research on backbone networks for medical image segmentation The core requirement of backbone networks for medical image segmentation is to adapt to the structural characteristics of different organs (such as small targets, large sizes, thin-layer high contrast, etc.) and multi-modal image differences 12 – 14 . Research in this field has carried out diversified explorations around structural innovation and performance balance, covering various classic and new architectures 15 . In the optimization of traditional CNN backbone networks, in addition to improvements targeting U-ResNet, scholars have also conducted in-depth research on U-Net series variants: some studies proposed U-Net++, which enhances the fusion of shallow details and deep semantic features by introducing a nested dense connection structure, improving the recognition ability of small lesions in the segmentation of abdominal organs such as the liver and kidneys 16 , 17 . However, the number of parameters of this architecture increases significantly, leading to insufficient adaptability to low-computing-power devices. Other studies designed lightweight U-Net (Mobile-U-Net), which uses depthwise separable convolution instead of standard convolution to compress the model size while maintaining basic accuracy, making it suitable for mobile ultrasound image segmentation 18 . Nevertheless, in high-resolution CT images, the limited receptive field results in reduced segmentation accuracy at the edges of large-sized organs. Research on 3D backbone networks is carried out to meet the segmentation needs of volumetric images: V-Net, as a classic 3D architecture, captures spatial three-dimensional features through a symmetric 3D encoder-decoder structure, and performs well in the segmentation of 3D structures such as lung nodules and vertebral bodies 19 , 20 . However, the computational complexity of 3D convolution is extremely high, and training can only be completed by relying on GPU parallel computing. To solve this problem, some scholars proposed an improved version of 3D-UNet, which introduces sparse convolution and attention gating mechanisms to reduce invalid computations while enhancing the features of key regions 21 , 22 . However, the segmentation integrity of thin-layer structures still needs to be improved. In the field of hybrid backbone networks, in addition to the combination of Transformer and CNN, pure Transformer architectures based on Vision Transformer (ViT) have also been applied to medical segmentation: some studies combined ViT with a decoder, and modeled the spatial correlation between organs through the global self-attention mechanism, realizing simultaneous recognition of multiple parts in whole-body bone scan image segmentation 23 , 24 . However, the segmentation accuracy of this model for small-sized structures (such as finger bones and cervical vertebral bodies) is lower than that of hybrid architectures, and its training relies on large-scale datasets. Comprehensively, existing research on backbone networks has formed a multi-dimensional system, but it still faces common challenges: the structural specificity of different organs leads to significant differences in the requirements for backbone networks, and there is a lack of universal and efficient architecture design; it is difficult to balance high accuracy and lightweight, and most models need to make trade-offs in specific scenarios; the adaptability to multi-modal and multi-site hybrid segmentation tasks is insufficient, and the cross-organ migration ability is limited. Attention mechanisms in medical image segmentation As a core technology to enhance the features of key regions, attention mechanisms in medical image segmentation have shifted from the application of general modules to organ-specific and pathology-adaptive designs, covering multiple dimensions such as channel, space, and shape 25 . In addition to explorations in spine segmentation, research on channel attention has also been widely applied in other fields: some scholars combined SENet with ResNet and applied it to breast X-ray image segmentation, strengthening the feature difference between calcifications and normal tissues through channel weight adjustment 26 . However, this module is sensitive to image noise and prone to weight allocation deviations in low-dose CT images. Another study proposed Efficient Channel Attention (ECA-Net), which simplifies the squeeze-and-excitation process of SENet and improves computational efficiency in the segmentation of fundus retinal blood vessels 27 . Yet, its attention to low-contrast structures such as thin blood vessels is insufficient. Research on spatial attention focuses on local region optimization: in brain image segmentation, some scholars designed a spatial attention module based on anatomical priors, which improves the segmentation accuracy of small structures such as the hippocampus and amygdala by introducing brain region anatomical atlases to constrain weight allocation 28 , 29 . However, this module relies on accurate prior information, and its applicability decreases in images with pathological deformation (such as brain region displacement caused by brain tumor compression). In skin lesion segmentation, researchers proposed multi-scale spatial attention, which captures the edge and internal texture features of lesions through receptive fields of different scales to improve the boundary blurring problem of nevi and hemangiomas 30 – 32 . However, multi-scale fusion increases the model complexity and affects the inference speed. As an emerging direction, shape-aware attention has also been applied in the segmentation of hard tissues such as bones and joints, in addition to spine segmentation: some studies constructed a bone shape dictionary to assist the shape-aware module in recognizing the morphological features of different bones, improving shape consistency in the segmentation of the pelvis and femur 33 – 35 . However, this method has weak adaptability to shape variations under pathological conditions such as fractures and deformities, and the construction of the dictionary requires a large amount of annotated data. The core limitations of current attention mechanism research are as follows: the matching degree between module design and organ characteristics is insufficient, and general modules are difficult to adapt to all scenarios; the collaborative fusion strategy of multi-attention dimensions (channel, space, shape) is not yet mature, and the optimization space of single-dimensional modules is limited; the robustness to pathological interference (such as tissue deformation caused by tumors and inflammation) is insufficient, and the segmentation stability is easily affected. Quantitative optimization and clinical connection research on loss functions for medical image segmentation As the guiding core of model training, the optimization research of loss functions has shifted from general strategies to medical scenario-specific designs, carrying out in-depth explorations around class imbalance, boundary accuracy, and clinical adaptation 36 . To address the class imbalance problem (such as low proportion of tumor regions and small lesion segmentation), in addition to weighted loss, scholars have proposed various improved schemes: Focal loss alleviates class bias in the segmentation of small targets such as lung nodules and tumors by reducing the weight of easy-to-classify samples and strengthening the learning of hard-to-classify regions 37 . However, this loss is sensitive to weight parameters and needs to be adjusted for different organs. Tversky loss optimizes the boundary misjudgment problem in the segmentation of cerebral hemorrhage and fundus lesions by introducing adjustable parameters to balance false positive and false negative errors 38 . Nevertheless, it is prone to gradient instability in high-noise images. The optimization of boundary loss focuses on clinical quantitative requirements: some studies proposed boundary loss based on distance transformation, which guides the model to optimize edge accuracy by calculating the Euclidean distance between the predicted boundary and the real boundary, improving continuity in the segmentation of tubular structures such as blood vessels and airways 39 . However, when this loss is used alone, it is easy to cause incomplete segmentation of the internal region. Other studies associated boundary loss with clinical indicators (such as tumor volume and organ deformation degree), and directly optimized the quantitative results through loss constraints, improving the accuracy of volume calculation in liver tumor segmentation 40 . However, this method needs to be customized for specific indicators, resulting in insufficient generality. In the research of combined loss functions, in addition to dynamic weight strategies, multi-loss collaborative optimization has become a new direction: some scholars proposed an adaptive selection mechanism for loss functions, which automatically switches the dominant loss according to the training stage (early regional segmentation, late boundary optimization), improving convergence efficiency in the segmentation of organs such as the heart and spleen 41 , 42 . However, the decision-making logic of this mechanism is complex, and a large number of experiments are required to verify the rationality of parameters. The key gaps in existing loss function research are as follows: the direct connection with clinical diagnosis needs is insufficient, most losses only optimize the similarity of segmentation masks, and are not associated with core clinical indicators (such as functional evaluation and treatment plan formulation); the generality across multi-modalities and multi-organs is poor, and redesign is required for different image types and organ characteristics; the robustness to annotation errors is insufficient, and in clinical weakly annotated datasets, loss guidance is prone to deviations. Methods Overview of the overall model architecture To achieve high-precision segmentation of multi-organ and multi-modal medical images while meeting clinical requirements for lightweight design and pathological adaptability, this study constructs an end-to-end model architecture. The core of the architecture consists of four modules working in collaboration: the data preprocessing module, the U-ResNet backbone network module, the shape-aware attention module, and the dynamically weighted combined loss function module. These modules are closely connected through data flow to form a complete segmentation pipeline. Input medical images undergo a series of data preprocessing steps, including grayscale normalization, region cropping, and data augmentation. This preprocessing pipeline eliminates device-induced differences and background interference, generating standardized samples suitable for model training or testing. These standardized samples are then fed into the U-ResNet backbone network. The U-ResNet adopts a symmetrical architecture to extract multi-scale features from the input images; its decoder module fuses shallow detail features and deep semantic features through skip connections, ensuring effective preservation of both fine-grained and high-level contextual information. Finally, a preliminary feature map containing target structural information is generated. The feature maps output by a specific layer of the decoder are passed to the shape-aware attention module. This module combines the contour prior features extracted in the preprocessing stage and strengthens target structure features while suppressing background interference through dynamic weight allocation, outputting optimized feature maps. The optimized feature maps are processed by subsequent layers of the decoder and the output layer to generate initial segmentation masks. The dynamically weighted combined loss function module adjusts the loss weights in real time based on the differences between the initial masks and the ground truth masks, combined with image features, guiding the model to optimize parameters through backpropagation and ultimately outputting segmentation results that meet clinical precision requirements. The overall model architecture can be clearly presented through an architectural diagram, corresponding to Fig. 1 . Fig. 1. Open in a new tab Overall architecture of the proposed model. Optimization design of U-ResNet backbone network Basic U-ResNet exhibits limitations in spinal segmentation, including fixed feature extraction that fails to adapt to varied spinal structure sizes, gradient instability in low-contrast regions, and inconsistent feature distributions across multi-modal images. To address these challenges, we optimize the network with targeted strategies while preserving its core U-shaped encoder-decoder structure and residual block advantages. The U-ResNet backbone adopts a U-shaped encoder-decoder design consisting of 4 encoder stages and 4 corresponding decoder stages: the encoder reduces feature map resolution via convolution and pooling layers to extract deep semantic features, while the decoder upsamples the feature map and fuses it with features from corresponding encoder layers through skip connections. Specifically, each encoder stage implements downsampling via 2 2 max-pooling with a stride of 2, and each decoder stage performs upsampling using transposed convolution with a kernel size of 2 2 and a stride of 2 to restore the feature map resolution to match the corresponding encoder stage. This fusion gradually reconstructs the feature map to generate final segmentation results. Residual blocks are systematically incorporated into all encoder and decoder stages of the U-ResNet: shallow layers (Encoder Stage 1/2, Decoder Stage 3/4) each contain 2 residual blocks to capture fine-grained anatomical details without excessive computational overhead, while deep layers (Encoder Stage 3/4, Decoder Stage 1/2) each integrate 3 residual blocks to enhance deep semantic feature extraction and mitigate gradient vanishing in complex spinal tissue regions. At the network’s core, residual blocks utilize identity mapping to prevent gradient vanishing during training, ensuring robust feature extraction even in deeper layers. The architecture of the U-ResNet backbone is illustrated in Fig. 2 . Fig. 2. Open in a new tab U-ResNet backbone network architecture. The mapping of traditional residual blocks is defined as: 1 where denotes input features, represents residual mapping, and is the output feature. To overcome the poor adaptability of fixed feature concatenation to varied spinal structure sizes, we replace simple concatenation with weighted sum-based fusion of shallow and deep features. This optimized skip connection strategy abandons the conventional direct concatenation method and adopts adaptive weighted fusion to align shallow and deep feature dimensions and enhance feature complementarity. The fusion formula is: 2 where and are shallow detail features and deep semantic features at spatial position , respectively, and is a dynamically adjusted weight based on spatial position. This fusion strategy emphasizes edge details for small spinal structures (e.g., intervertebral discs) and captures global features for large ones (e.g., vertebral bodies), thereby enhancing adaptability to size variations. To resolve the mismatch between fixed convolutional receptive fields and diverse spinal target sizes, we introduce a dynamic receptive field convolution that adapts kernel size to the target structure. Its calculation is given by: 3 where is the convolution output at position , is the input feature map value at position , denotes the convolution kernel, and is the receptive field–adaptively adjusted according to target size. This allows the network to focus on local features for small targets and capture global context for large ones. To mitigate gradient instability in low-contrast regions, we integrate a gradient smoothing term into the residual mapping, which is formulated as: 4 where is the original residual mapping, is the smoothing coefficient, and represents the gradient operator. This smoothing effectively alleviates gradient vanishing, particularly in regions with weak contrast such as tissue boundaries. To unify inconsistent feature distributions across multi-modal spinal images, we propose modality-adaptive normalization, which standardizes features by computing the mean and variance of the feature map over the spatial domain. This normalization module is embedded immediately after the residual block output in each encoder/decoder stage to ensure cross-modal feature alignment throughout the feature extraction process. The normalization formula is: 5 where and are the mean and standard deviation of the feature map across the spatial domain. This process ensures consistent feature distributions across different modalities (MRI/CT), improving the model’s adaptability in cross-modal segmentation tasks. Design of shape-aware attention module The human spine exhibits highly stereotyped anatomical shape characteristics that are critical for accurate segmentation: vertebral bodies possess a regular elliptical cortical bone structure with distinct contour boundaries, and intervertebral discs present a circular soft tissue morphology with a characteristic thickness and fixed spatial arrangement relative to vertebral bodies. However, common spinal diseases (e.g., intervertebral disc herniation, vertebral compression fractures, scoliosis) induce significant shape deformations–vertebral bodies may show irregular compression/fracture morphology, intervertebral discs exhibit bulging/herniation or degenerative shrinkage, and scoliosis causes abnormal curvature of the spinal column, destroying the inherent anatomical shape regularity. Conventional attention mechanisms focus only on channel/spatial feature correlations without incorporating shape priors, leading to attention bias to background tissues or mis-segmentation of deformed spinal structures, which severely degrades segmentation accuracy for pathological spinal images. This anatomical specificity and disease-induced shape variation of the spine constitute the core motivation for designing the shape-aware attention module: it leverages spinal anatomical shape priors to enhance target discrimination and adaptively adjusts to disease-induced deformations, addressing the unique segmentation challenges of spinal disease images that conventional attention mechanisms cannot resolve. In spinal image segmentation, target shape variations and background interference degrade segmentation accuracy. To address these issues, we develop a shape-aware attention module that enhances target shape features and spatial correlations through feature fusion and dynamic weight allocation–this module is integrated at the feature fusion stage between each encoder and corresponding decoder stage of the U-ResNet architecture, and operates on both the deep semantic features output by the encoder and the high-resolution shallow features output by the decoder. It takes semantic features from the decoder and contour prior features from preprocessing as inputs, and aligns their dimensions before feeding them into a feature interaction layer to enable subsequent attention computation. The core operational principle of the shape-aware attention mechanism is to extract shape information from contour prior features and semantic feature correlations, then encode this shape information into adaptive attention weights to modulate feature weighting: first, contour prior features C ( s ) are pre-extracted from spinal images using edge detection and anatomical prior constraints, which encode the inherent shape characteristics of normal spinal structures (e.g., the elliptical shape of vertebral bodies and the circular shape of intervertebral discs); semantic features S ( s ) integrate spatial and contextual information of spinal targets from encoder-decoder fusion features, reflecting the actual shape of spinal structures in the input image. To measure the local correlation strength between semantic and contour prior features at each spatial point, we use a mutual information density function. Its calculation is: 6 where S ( s ) denotes semantic features (encoding actual shape information of spinal structures at spatial point s ) and C ( s ) represents contour prior features (encoding anatomical shape priors of spinal structures at spatial point s ). This mutual information density function quantifies the consistency between the actual shape (from semantic features) and the prior shape (from contour priors) at each spatial point, serving as the fundamental metric for extracting shape-related feature correlation information. We then convert this correlation into spatial attention weights to highlight regions with strong alignment between semantic and shape features. The initial attention weight function is: 7 Here, the numerator captures the local shape consistency (correlation) at point s , and the exponential function amplifies the difference in correlation values to enhance the distinguishability of shape-matched regions; the denominator normalizes results across the entire spatial domain –ensuring falls within the range [0, 1], which converts the shape consistency metric into initial spatial attention weights: higher values indicate stronger alignment between the actual shape and prior shape at spatial point s , meaning the region is more likely to be a spinal target and should receive higher feature weighting. To account for shape variations under pathological conditions, we introduce a dynamic shape adaptation factor that quantifies the deviation between the target’s actual shape and prior contour. This factor is defined as: 8 where is the actual contour feature of the target structure (obtained via edge detection on the input image, directly reflecting the real shape of pathological spinal structures such as deformed vertebral bodies or herniated intervertebral discs). The term quantifies the shape deviation between the prior contour and actual contour at spatial point s ; dividing by normalizes the deviation to the range [0, 1], and subtracting from 1 ensures that decreases as the target shape deviates more from the prior contour (with ). A lower reduces the module’s reliance on prior contour information, which is critical for adapting to pathological shape variations in spinal structures (e.g., compressed vertebral bodies or herniated intervertebral discs caused by spinal diseases). We further combine the dynamic adaptation factor with the initial attention weight to generate final spatial attention weights, ensuring smooth adaptation to shape changes. The final weight function is: 9 This formula modulates the initial attention weight (based on prior shape) with the dynamic shape adaptation factor : when is high (small shape deviation, e.g., normal spinal structures), the module mainly relies on to weight features, prioritizing regions consistent with anatomical shape priors; when decreases (indicating significant shape variation, e.g., pathological spinal structures such as scoliosis-induced curved vertebrae), the model increases reliance on the average attention weight of local neighborhood features , avoiding attention bias caused by failed contour priors and ensuring adaptive feature weighting for abnormal spinal shapes induced by diseases. To generate optimized feature maps, we apply the computed attention weights to input features–enhancing target features while suppressing background interference. The feature optimization formula is: 10 where F ( s ) is the input feature at spatial point s (fused encoder-decoder features containing shape and semantic information). The first term directly enhances the feature representation of target spinal regions by multiplying with the shape-guided attention weight A ( s )–the higher the A ( s ) value (stronger shape consistency), the greater the feature enhancement; the second term uses a Gaussian convolution kernel to smooth and suppress background features (regions with low A ( u ) values), where inversely weights background features based on shape information to reduce their interference. This two-term design ensures that shape information encoded in A ( s ) directly drives differential feature weighting: spinal target regions with consistent shape characteristics (even deformed ones caused by diseases) receive enhanced feature representation, while background regions and abnormal shape regions are adaptively adjusted to maintain segmentation accuracy for spinal disease images. Design of dynamically weighted combined loss function Medical image segmentation for spinal tasks faces three key challenges: class imbalance, boundary blurring, and the need for clinical quantification. To address these, we develop a dynamically weighted combined loss function that integrates multiple loss components with feature-adaptive weights, balancing global region segmentation accuracy and local boundary clarity while meeting clinical quantitative demands. To tackle class imbalance and ensure sufficient learning for small lesions, we define a regional loss function with density-based adaptive weights: 11 where is the ground truth mask, is the predicted probability, and weight factor adjusts based on target region density–directing the model to focus more on small lesion regions. To enhance boundary segmentation clarity, which is critical for distinguishing vertebrae and intervertebral discs, we introduce a boundary loss function that prioritizes boundary pixels via distance-based weighting. Its calculation is: 12 where is the boundary distance weight, ensuring the model allocates more learning focus to boundary pixels and improves boundary segmentation precision. To balance the contributions of regional and boundary losses across diverse segmentation scenarios, we dynamically adjust weights for regional loss and for boundary loss based on target region proportion and boundary clarity. The adjustment formula for is: 13 This strategy prioritizes regional loss when segmenting small targets and shifts focus to boundary loss when dealing with blurry boundaries, adapting to task-specific needs. The total loss integrating regional and boundary components is a weighted sum: 14 For images containing small lesions, increases to emphasize regional loss and ensure complete small-target segmentation; for images with blurry boundaries, rises to enhance boundary loss and refine edge clarity. To meet clinical quantification requirements such as accurate vertebral or disc volume calculation, we add a volume constraint loss that ensures the model outputs reliable quantitative results for clinical evaluation. This loss is formulated as: 15 where is the ground truth volume of the target structure. We further integrate the volume constraint loss into the total loss to form the final loss function with clinical constraints: 16 This design simultaneously optimizes segmentation accuracy via regional and boundary losses and clinical quantification reliability via volume constraint loss, providing a comprehensive solution for spinal medical image segmentation. Experiment Dataset description Lumbar Spine MRI Dataset comprises 200 T2-weighted MRI scans of the lumbar spine from 200 distinct patients, 108 males and 92 females, with age range 25–78 years and mean age 49.6 years. The original image resolution is 512 512 pixels with a slice thickness of 3 mm. The dataset includes diverse spinal conditions with clear category distribution: 45 normal cases accounting for 22.5 percent, 78 intervertebral disc herniation cases accounting for 39.0 percent, 57 intervertebral disc degeneration cases accounting for 28.5 percent, and 20 vertebral osteophytes cases accounting for 10.0 percent. The dataset offers heterogeneous data covering different degrees of lumbar spine pathologies, facilitating model training and enabling identification and classification of lumbar spine abnormalities. VerSe Dataset consists of 194 whole-spine CT scans from 194 patients, 87 males and 107 females, with age range 18 to 85 years and mean age 56.2 years. The original voxel resolution is 0.625 0.625 1.0 mm 3 . The dataset covers cervical, thoracic, lumbar, and sacral spine regions, with the following disease category distribution: 86 normal spine cases accounting for 44.3 percent, 62 scoliosis cases accounting for 32.0 percent including mild 28 cases, moderate 22 cases and severe 12 cases, and 46 vertebral compression fracture cases accounting for 23.7 percent including mild compression 18 cases, moderate compression 20 cases and severe compression or fracture 8 cases. The dataset provides full representation of healthy and pathological spinal structures, ideal for model training and whole-spine segmentation. Preprocessing All spinal MRI and CT images were preprocessed following a standardized pipeline tailored to each modality, including signal correction, spatial alignment, region cropping, normalization, and data augmentation, to ensure sample consistency and improve model robustness. Additionally, a dedicated preprocessing sub-pipeline was implemented to extract contour prior features for the shape-aware attention module, with detailed steps as follows: Contour prior feature extraction: (1) Edge detection: After image normalization (min-max normalization for MRI/CT as described below), the Canny edge detector was applied with adaptive thresholding (low threshold = 0.1 max grayscale value, high threshold = 0.3 max grayscale value) to initial contour extraction, and the Gaussian smoothing kernel ( = 1.5) was used to reduce noise interference before edge detection. (2) Morphological refinement: Binary morphological operations (dilation with structuring element followed by erosion) were performed to connect discontinuous contour segments and remove isolated noise points. (3) Contour feature encoding: The refined contour maps were converted to distance maps (Euclidean distance transform) to quantify the spatial relationship between each pixel and the nearest contour, with distance values normalized to 0–1. (4) Feature fusion preparation: The contour distance maps were resized to match the spatial resolution of the U-ResNet backbone input ( pixels) via bilinear interpolation, and concatenated with the original normalized images as multi-channel input to the shape-aware attention module. Lumbar spine MRI images: (1) Signal correction: The N4ITK method was applied to fix signal unevenness caused by magnetic field inhomogeneity with verification error less than 0.05. (2) Spatial alignment: Aligned to the sagittal plane using vertebral body centroids to eliminate spatial offset. (3) Region cropping: Cropped to L1–L5 vertebral segments with size pixels to remove irrelevant background. (4) Resizing: Resized to pixels via bilinear interpolation. (5) Normalization: Grayscale values normalized to the range of 0–1 using min-max normalization to eliminate intensity differences. (6) Data augmentation: Spatial augmentation including random rotation , scaling 0.8–1.2 times, and translation pixels and pixel-level augmentation including Gaussian noise and contrast adjustment were applied; oversampling was used for rare pathologies to balance sample distribution. VerSe whole-spine CT images: (1) HU value adjustment: Window width 1500 HU and window level 300 HU were applied to enhance tissue contrast, with metal artifact reduction to minimize streak artifacts. (2) Spatial alignment: Aligned to the coronal plane using spinal midline detection. (3) Slicing and cropping: 3D volumes sliced into 2D images with 1 mm thickness and cropped to pixels covering the spinal column. 4) Normalization: HU values clipped to the range of -1000 to 1000 and normalized to the range of 0 to 1 to eliminate density differences. 5) Data augmentation: Spatial augmentation including random rotation , scaling 0.9 to 1.1 times, and translation pixels and pixel-level augmentation including HU shift and Poisson noise were applied; oversampling increased rare lesions to 15 percent of the training set. Annotation process: All spinal image annotations include segmentation masks for vertebral bodies and intervertebral discs, as well as corresponding disease category labels for pathological regions. A two-round annotation check was adopted to ensure accuracy, and the inter-annotation consistency was quantified using Cohen’s Kappa coefficient with average value 0.88 for categorical labels and Dice Similarity Coefficient with average value 0.92 for segmentation masks, verifying high reliability of the annotation results. Experiment details Experiments ran on dual RTX 4090 GPU setup with 128 GB DDR5 RAM. Handled computational needs for training deep learning models on large medical image datasets. Implemented with Python 3.8 TensorFlow and PyTorch as main frameworks. The Adam optimizer was adopted as the optimization strategy with = 0.9, = 0.999 and weight decay = to balance convergence speed and stability. A cosine annealing learning rate scheduler was used, with an initial learning rate of , minimum learning rate of , and set to 50 (matching total training epochs) to adaptively adjust the learning rate during training. Batch size set to 8 for Lumbar Spine MRI dataset 6 for VerSe dataset based on memory limits. The initial learning rate was set to for both datasets, and the learning rate was reduced by 10% every 10 epochs before switching to cosine annealing after epoch 20. Total training epochs set to 50. Convergence criteria were defined as follows: the model was considered converged when the validation Dice coefficient reached a plateau (variation < 0.001) for 3 consecutive epochs, or when the maximum number of epochs (50) was reached. Early stopping used stopped training if validation metrics didn’t get better for 5 straight epochs. Key hyperparameters of the shape-aware attention module were configured as follows: the kernel size of the convolution layer for local shape feature extraction was set to , while the kernel size for global shape context modeling was ; Layer Normalization was adopted as the normalization strategy, with a momentum of 0.9 and an epsilon of to stabilize the calculation of attention weights. Adapted to task needs dynamically weighted combined loss used for segmentation on Lumbar Spine MRI dataset. Cross-entropy loss used for classification. Weight ratio between segmentation and classification losses found through experiments. For VerSe dataset initial weights for regional and boundary losses set to 0.5. Adaptive weighting used for small vertebrae. Segmentation performance checked with metrics Dice coefficient and Hausdorff distance for Lumbar Spine MRI dataset. Average Dice used for VerSe dataset. Metrics compared every 5 epochs. Changes shown through TensorBoard. Full-spine segmentation accuracy was evaluated every 10 epochs for the VerSe dataset, and Dice distribution heatmaps were generated for accuracy tracking. Hierarchical training was employed, starting with training the backbone network for 20 epochs, followed by joint training with the attention module. A difficulty-adaptive sampling mechanism was introduced to improve segmentation accuracy, where the sampling probability for hard-to-segment samples was increased. To reduce overfitting, Dropout layers were added, and the Dropout rate was dynamically adjusted if the Dice coefficient decreased for two consecutive epochs, balancing overfitting suppression and model convergence. Baseline model selection Seven types of state-of-the-art (SOTA) models in the field of medical image segmentation since 2023 are selected as baselines, covering mainstream technical routes. Their core information is as follows: U-KAN 43 : Replaces the MLP layers in U-Net with KAN layers to enhance nonlinear feature capture. It excels in handling blurred boundaries (e.g., intervertebral discs), has low computational cost, and is suitable for spine soft tissue segmentation. TransUNet 44 : Optimizes the CNN-Transformer collaborative architecture, supporting global context extraction and candidate region refinement. It provides 2D/3D implementations and is suitable for long-sequence segmentation of the entire spine. KAN-Mamba FusionNet 45 : Integrates the nonlinear modeling capability of KAN and the long-range feature processing capability of Mamba, with a dedicated KAMBA module. It is suitable for the correlated segmentation of multiple spine structures and pathological region segmentation. CTRCL 46 : A CNN-Transformer collaborative learning framework that optimizes knowledge transfer through RLCL/CFCL modules. It balances local details and global dependencies, improving the segmentation accuracy of small/abnormal vertebrae. nnFormer 47 : A 3D Transformer model that introduces volumetric self-attention and skip attention. It features high computational efficiency, is suitable for full-spine 3D segmentation on the VerSe dataset, and excels in capturing spatial morphological features. Medical SAM 2 48 : Improved based on SAM 2, it converts segmentation into video tracking, supporting one-prompt multi-image segmentation. It optimizes the segmentation consistency of the entire spine and reduces cross-slice errors. VM-UNet++ 49 : An improved Vision Mamba architecture with added cross-scale attention to enhance fine-structure segmentation. It is suitable for segmenting small components of the spine (e.g., fracture lines, herniated regions) and has fast inference speed. SPINEPS 50 : A two-phase deep learning method dedicated to whole-spine segmentation in T2-weighted MR images, realizing semantic and instance segmentation of 14 spinal structures (including vertebra substructures, intervertebral discs, spinal cord, etc.). It outperforms nnUNet baseline in boundary precision and multi-structure segmentation consistency, and is the first publicly available algorithm for this task. 3D MFA 51 : A 3D Multi-Feature Attention model optimized with RCBAM and Feature Pyramid Pooling, trained separately on multi-view 3D spine images. It achieves high Dice coefficient and IoU on VerSe datasets with low computational cost, excelling in handling unclear boundaries and tissue overlap in spine segmentation. All baseline models share the same training configuration as the proposed model: AdamW optimizer (weight decay of ), initial learning rate of (cosine annealing scheduling), and batch size of 8. A unified preprocessing pipeline is adopted, and all models are trained to convergence in the same experimental environment (NVIDIA RTX 4090 GPU, PyTorch 2.0). Evaluation metrics For the spine segmentation and disease classification tasks, corresponding evaluation metrics are set respectively: All selected metrics are specifically tailored to the anatomical characteristics of spinal structures and the clinical requirements for structural accuracy in medical image segmentation, ensuring each metric directly reflects the clinical utility and structural precision of segmentation and classification results for spinal regions. The segmentation task adopts the Dice Similarity Coefficient (DSC) to measure the overlap between the segmentation results and the ground truth. DSC is critical for spinal segmentation as it quantifies the volumetric overlap of key spinal anatomical structures including vertebral bodies and intervertebral discs. Spinal structures such as intervertebral discs have small physical dimensions and require high overlap accuracy to ensure clinical reliability. DSC directly reflects whether the model can accurately capture the complete range of these small target structures without over-segmentation or under-segmentation, both of which would compromise the clinical value of the segmentation results. The segmentation task also uses the 95% Hausdorff Distance (95% HD) to quantify the maximum distance between the segmentation boundary and the ground truth boundary. 95% HD is of paramount importance for assessing the structural accuracy of spinal pathological regions. This metric focuses on extreme boundary deviations and is essential for identifying spinal pathologies that depend on precise boundary localization. Spinal pathologies including vertebral compression fractures with irregular cortical bone boundaries, scoliosis with abnormal spinal curvature and intervertebral disc herniation with protruded soft tissue boundaries all require accurate boundary segmentation. A small 95% HD value ensures the model’s segmentation boundaries align with true anatomical boundaries and avoids misdiagnosis risks caused by boundary shifts in clinical applications. Additionally, the Mean Surface Distance (MSD) is used to calculate the average distance between the segmentation boundary and the ground truth boundary. MSD complements 95% HD by providing a global assessment of boundary consistency across the entire spinal structure. It reflects the overall structural accuracy of segmented vertebral segments rather than just extreme deviations. This is vital for evaluating the model’s ability to maintain consistent segmentation quality across the entire spine including cervical, thoracic and lumbar regions which have varying morphological characteristics. The classification task uses Accuracy (Acc) to measure the overall classification ability of the model. Accuracy reflects the model’s general ability to distinguish normal from pathological spinal conditions and aligns with clinical needs for high overall diagnostic accuracy in spinal disease screening. It provides a foundational assessment of the model’s performance in identifying spinal pathologies at a population level. The classification task also adopts Precision (P) and Recall (R) to adapt to imbalanced sample classification. Precision and Recall address the challenge of imbalanced spinal disease sample distributions. Rare pathological types such as vertebral burst fractures often account for a small proportion of clinical datasets. Precision ensures the model minimizes false positives in disease diagnosis which is critical for avoiding unnecessary clinical interventions. Recall guarantees the model does not miss rare but clinically significant spinal pathologies, making both metrics indispensable for reliable clinical decision-making. Moreover, the Kappa coefficient is used to evaluate the consistency between the classification results and the ground truth, which is applied to ordinal classification tasks such as intervertebral disc degeneration grading. The Kappa coefficient goes beyond simple accuracy by accounting for chance agreement and is uniquely suitable for ordinal classification tasks such as intervertebral disc degeneration grading. It quantifies the model’s ability to correctly distinguish between different severity levels of spinal pathologies. This ability directly impacts clinical treatment planning as different severity levels of spinal pathologies require different intervention strategies. The Kappa coefficient therefore reflects the structural accuracy of the model’s pathological severity assessment for spinal regions. Results Experimental results and analysis of overall model performance From the table 1 , the proposed “Ours” model in this study achieves the optimal performance across all core metrics on both datasets, with excellent performance stability. Our model performance fluctuation standard deviation mostly 0.6–0.8 range, same level as top models, lower than U-KAN, nnFormer and other models, showing good adaptability and reliability across different spinal image data. On the Lumbar Spine MRI dataset, our model achieves an accuracy of 86.7 ± 0.6, which is higher than all other comparative models, including 85.3 ± 0.6 of 3D MFA, 85.1 ± 0.6 of CTRCL, 84.8 ± 0.7 of SPINEPS, 84.5 ± 0.7 of Medical SAM 2, and 84.2 ± 0.7 of KAN-Mamba FusionNet. The F1 score of our model is 84.2 ± 0.7, which is also superior to all comparative models, surpassing 3D MFA (83.2 ± 0.7), CTRCL (82.9 ± 0.7), SPINEPS (82.7 ± 0.7), Medical SAM 2 (82.3 ± 0.7) and KAN-Mamba FusionNet (82.0 ± 0.7). This advantage stems from the customized optimizations of the U-ResNet backbone network, the integrated shape-aware attention module, and the applied dynamic combination loss function, all of which help improve the completeness and accuracy of small object segmentation. On the VerSe dataset, the performance advantage of our model is more obvious, with an accuracy of 87.9 ± 0.6, which is higher than all comparative models, such as 86.5 ± 0.6 of 3D MFA, 86.3 ± 0.6 of CTRCL, 85.9 ± 0.7 of SPINEPS, 85.6 ± 0.7 of Medical SAM 2, and 85.1 ± 0.7 of KAN-Mamba FusionNet. The F1 score is 85.8 ± 0.6, which is also better than all other models, exceeding 3D MFA (84.4 ± 0.6), CTRCL (84.2 ± 0.6), SPINEPS (83.8 ± 0.7), Medical SAM 2 (83.5 ± 0.7) and KAN-Mamba FusionNet (83.0 ± 0.7). The good performance indicates that our model can well adapt to the morphological variations of different spinal segments. The lightweight U-ResNet architecture maintains segmentation accuracy while avoiding generalization degradation caused by excessive parameters. Compared with other models, U-KAN and nnFormer perform well in non-spine segmentation tasks but have lower performance on spinal images due to the lack of optimization for spinal anatomy, with F1 scores all below 82. Medical SAM 2 is a general segmentation model with certain cross-scene adaptability but insufficient segmentation accuracy for spinal pathological regions, whose F1 score is 2–3 percentage points lower than that of our model. KAN-Mamba FusionNet integrates nonlinear pattern modeling and long-range dependency modeling, but its network structure is too complex, making it prone to overfitting on small-sample spinal data, and its performance stability (standard deviation 0.7–0.8) is slightly worse than that of our model. The lower standard deviation of our model further indicates its reduced sensitivity to image noise, and the high precision and recall values reflect its effectiveness in mitigating boundary leakage and mis-segmentation of adjacent spinal structures compared to other models. Table 1. Performance comparison of different models on lumbar spine MRI and VerSe datasets. Model Lumbar spine MRI dataset VerSe dataset Acc P R F1 Acc P R F1 U-KAN TransUNet KAN-Mamba FusionNet CTRCL nnFormer Medical SAM 2 VM-UNet++ SPINEPS 3D MFA Ours Open in a new tab From the perspective of the Kappa coefficient performance for intervertebral disc degeneration grading on the Lumbar Spine MRI Dataset in Table 2 , the Kappa coefficients of all models range from 0.76 to 0.83 with a consistent standard deviation of 0.02. This indicates that the scoring stability of each model on this task is similar, and there is no significant performance fluctuation. The proposed “Ours” model achieves the highest Kappa coefficient of 0.83 ± 0.02, which is significantly superior to other comparison models, demonstrating better diagnostic consistency in the intervertebral disc degeneration grading task. nnFormer has the lowest Kappa coefficient of 0.76 ± 0.02, which is consistent with its characteristic of not being optimized for anatomical structures in spinal image segmentation tasks. U-KAN and VM-UNet++ show similar performance, with both having a Kappa coefficient of 0.77 ± 0.02. Although these two models have a certain foundation in segmentation accuracy, they are insufficient in the “degree differentiation” ability required for grading tasks and cannot effectively distinguish the characteristic differences between mild, moderate, and severe degeneration. The Kappa coefficients of TransUNet, KAN-Mamba FusionNet and SPINEPS are 0.78 ± 0.02, 0.79 ± 0.02 and 0.79 ± 0.02 respectively, which are slightly higher than those of the previous three models. According to the data in Table 1 , KAN-Mamba FusionNet achieves better overall feature capture of degenerated regions due to the integration of nonlinear pattern modeling and long-range dependency modeling capabilities, but it still fails to reach a high level of consistency. CTRCL, Medical SAM 2 and 3D MFA perform relatively well, all with a Kappa coefficient of 0.80 ± 0.02. CTRCL improves feature discrimination through the collaborative learning of CNN and Transformer, while Medical SAM 2 achieves a certain grading ability relying on the cross-scenario adaptability of general segmentation. However, all these models still have the possibility of misjudgment when facing “mild degeneration with blurred boundaries” and “moderate degeneration with atypical local features” in degeneration grading. Based on the performance of each model in Table 2 , the core reason why the Ours model can achieve a high Kappa coefficient of 0.83 ± 0.02 lies in its collaborative architecture designed for spinal images: the customized optimization of the U-ResNet backbone ensures the extraction of subtle features in degenerated regions; the shape-aware attention module accurately focuses on the key regions of intervertebral disc degeneration; and the dynamic combined loss function strengthens the learning of category differences among different degeneration degrees. The joint effect of these three components enables the model to achieve higher diagnostic consistency in the grading task, which is more in line with the clinical requirements for the accuracy of intervertebral disc degeneration assessment. Our model’s higher Kappa coefficient also implies its reduced mis-segmentation risk for degenerated discs with blurred boundaries, and the consistent standard deviation confirms its insensitivity to image noise in grading tasks, avoiding boundary leakage-induced misjudgment of degeneration severity. Table 2. Comparison of Kappa coefficients for intervertebral disc degeneration grading on the lumbar spine MRI dataset. Model Kappa coefficient U-KAN TransUNet KAN-Mamba FusionNet CTRCL nnFormer Medical SAM 2 VM-UNet++ SPINEPS 3D MFA Ours Open in a new tab Table 3 and Table 4 results show our proposed model (“Ours”) achieves the best performance on all core metrics, with good adaptability in spinal image segmentation tasks. On the Lumbar Spine MRI dataset, the Ours model obtains a vertebral body Dice Similarity Coefficient (DSC) of 0.899 ± 0.010, surpassing the second-best 3D MFA by 0.010 and CTRCL by 0.012. Its intervertebral disc DSC reaches 0.841 ± 0.013, 0.011 higher than 3D MFA and 0.013 higher than CTRCL. These improvements mean the Ours model not only works well on regular vertebral bodies but also excels at capturing blurry boundary regions like intervertebral discs. In terms of boundary consistency, the 95% Hausdorff Distance (HD) and Mean Surface Distance (MSD) of the Ours model are 2.65  ±  0.08 mm and 0.78  ±  0.05 mm, outperforming 3D MFA by 0.23 mm and 0.07 mm, and CTRCL by 0.26 mm and 0.09 mm, which shows the Ours model has better boundary segmentation performance, stemming from the shape-aware attention module that enhanced boundary feature learning. On the VerSe dataset, the advantages of the Ours model are more obvious. Its small vertebrae DSC is 0.835 ± 0.014, 0.12 higher than the second-best 3D MFA and 0.014 higher than CTRCL, and the abnormal vertebrae DSC is 0.857 ± 0.012, 0.012 higher than 3D MFA and 0.014 higher than CTRCL. These results demonstrate the model’s strong robustness in segmenting vertebral bodies of different sizes and pathological states. The 95% HD and MSD of the Ours model are 2.82 ± 0.09 mm and 0.83 ± 0.05 mm, outperforming 3D MFA by 0.30 mm and 0.07 mm, and CTRCL by 0.33 mm and 0.09 mm, which proves its ability to maintain good boundary consistency even in complex whole spine segmentation tasks. Compared with other models, the Ours model consistently achieves higher performance, especially in handling the boundary segmentation of small and abnormal vertebrae. U-KAN and nnFormer show lower performance due to the lack of spinal anatomy optimization, and they are less effective in segmenting finer structures such as intervertebral discs and small vertebrae. Medical SAM 2 has good adaptability in general segmentation tasks but still lags behind the Ours model in spinal pathology segmentation. KAN-Mamba FusionNet adopts complex network designs for long-range dependency modeling but cannot match the Ours model in segmentation accuracy and boundary control for small vertebrae. The superior segmentation accuracy and boundary consistency of the Ours model benefit from the synergistic effect of the customized U-ResNet backbone, shape-aware attention module and dynamic combined loss function, which makes it outperform all other comparison models in both local lumbar spine segmentation and entire spine multi-type vertebral body segmentation, and fully meets the clinical requirements for high precision and stability in spinal image segmentation. Table 3. Comparison of lumbar spine segmentation results on the lumbar spine MRI dataset. Model Vertebral body DSC Intervertebral Disc DSC 95% HD (mm) MSD (mm) U-KAN TransUNet KAN-Mamba FusionNet CTRCL nnFormer Medical SAM 2 VM-UNet++ SPINEPS 3D MFA Ours Open in a new tab Table 4. Comparison of whole spine segmentation results on the VerSe dataset. Model Small vertebrae DSC Normal vertebrae DSC Abnormal vertebrae DSC 95% HD (mm) MSD (mm) U-KAN TransUNet KAN-Mamba FusionNet CTRCL nnFormer Medical SAM 2 VM-UNet++ SPINEPS 3D MFA Ours Open in a new tab From Table 5 , the proposed model achieves the optimal computational efficiency on both datasets while maintaining the highest segmentation performance. It has the lowest FLOPs (26.3 G) and parameters (14.5 M) among all comparison models, with the shortest inference time of ms/sample on the Lumbar Spine MRI Dataset and ms/sample on the VerSe Dataset. The inference time of all models on the VerSe Dataset is slightly higher than that on the Lumbar Spine MRI Dataset, which is attributed to the more complex morphological characteristics and larger image scale of the whole spine samples in the VerSe Dataset. Even so, the proposed model still maintains the minimum inference time and the smallest inference time increase between the two datasets, showing strong adaptability. Compared with other models, complex models such as KAN-Mamba FusionNet and Medical SAM 2 have significantly higher FLOPs and inference time (exceeding 90 ms/sample on the VerSe Dataset) due to their intricate network structures. Traditional models and top-performing segmentation models (e.g., 3D MFA, CTRCL) also lag behind the proposed model in computational efficiency. The superior computational efficiency of the proposed model benefits from the lightweight optimization of the U-ResNet backbone and the sparse computing strategy of the shape-aware attention module, which effectively reduces computational burden. Its inference time of less than 65 ms/sample on both datasets meets clinical real-time processing requirements, and the small number of parameters facilitates deployment on low-power medical edge devices, verifying its feasibility for clinical deployment. Table 5. Computational complexity and inference time comparison of different models on two datasets. Model FLOPs (G) Params (M) Inference time (ms/sample) Lumbar spine MRI dataset VerSe dataset U-KAN 28.6 15.8 TransUNet 32.4 18.3 KAN-Mamba FusionNet 45.7 22.6 CTRCL 30.2 17.5 nnFormer 36.9 20.1 Medical SAM 2 41.5 24.8 VM-UNet++ 29.8 16.7 SPINEPS 31.7 18.0 3D MFA 33.5 19.2 Ours Open in a new tab Figure 3 presents the overall loss curves of the model on the datasets, including the loss changes of both the training set and the validation set. From the trend of the curves, as the number of training epochs increases, the loss of all datasets shows a decreasing trend and gradually stabilizes after approximately 10 epochs. This indicates that the model continuously learns and converges effectively during the training process. The training loss (SPIDER-Train) and validation loss (SPIDER-Val) of the SPIDER dataset are generally lower than those of the VerSe dataset. This reflects that the sample or task characteristics of the SPIDER dataset are more suitable for model training, enabling the model to achieve a more significant loss reduction on this dataset and eventually stabilize at a lower level. Meanwhile, the gap between the loss curves of the training set and the validation set is small, which indicates that the model does not exhibit obvious overfitting and maintains good generalization ability on both datasets. It can stably learn features from the training data and transfer them to the segmentation tasks of the validation data. This provides support for the high-precision performance of the model in spinal image segmentation from the perspective of the training process. Fig. 3. Open in a new tab Overall loss curves on the SPIDER and VerSe datasets. Ablation experiment results and analysis The ablation experiment results in Table 6 clearly demonstrate the independent impact of the U-ResNet backbone and shape-aware attention module on the model’s segmentation and classification performance, and verify the indispensable role of each core module. Removing U-ResNet leads to the most significant performance degradation on both datasets: the accuracy of the Lumbar Spine MRI Dataset drops by 3.2 percentage points, and that of the VerSe Dataset drops by 3.1 percentage points, with the F1 score decreasing by 3.0 and 3.1 percentage points respectively. This fully proves that the U-ResNet backbone, as the core feature extraction module of the model, has a unique independent contribution to the performance improvement: its dynamically adjustable receptive field can better adapt to the scale differences of different vertebral segments in the lumbar spine and the entire spine, effectively extract fine-grained anatomical and pathological feature information of spinal structures, and provide a more accurate and robust feature foundation for subsequent classification and segmentation tasks. Without this module, the model’s ability to capture key spinal features is significantly weakened, leading to a noticeable decline in overall performance. Removing the shape-aware attention module results in a decrease of about 2.3 percentage points in the F1 score on both datasets, with the accuracy of the Lumbar Spine MRI Dataset dropping by 2.5 percentage points and that of the VerSe Dataset by 2.4 percentage points. This verifies the independent positive impact of the shape-aware attention module on the model’s performance: the module can independently enhance the model’s feature discrimination ability by adaptively encoding spinal shape information, effectively focus the model’s attention on key pathological and abnormal spinal regions such as degenerated intervertebral discs, herniated discs and scoliotic vertebrae, and reduce the interference of irrelevant background soft tissues such as muscles and fat on feature learning. When this module is removed, the model’s attention distribution becomes scattered, the feature discrimination ability for spinal target regions is reduced, and the classification and segmentation accuracy are negatively affected. Removing the dynamic weighted combined loss causes the accuracy of the Lumbar Spine MRI Dataset to drop by 2.9 percentage points and that of the VerSe Dataset by 2.8 percentage points, with a corresponding decline in the F1 score. This loss function can effectively balance the signal differences between vertebral bodies and intervertebral discs during training, ensure the accurate learning of features for smaller spinal structures such as intervertebral discs, and address the category imbalance problem between normal and abnormal vertebrae, thus improving the model’s ability to identify and segment various pathological types of vertebrae. Without this module, the learning weights of different spinal targets are imbalanced, leading to a decline in the accuracy and integrity of segmentation and classification. The full version of the model achieves the optimal performance on all evaluation metrics of both datasets, which fully demonstrates the excellent synergistic effect between the U-ResNet backbone and the shape-aware attention module: the U-ResNet backbone provides high-quality fine-grained feature information for the model, and the shape-aware attention module further mines and enhances the shape-related feature information in the extracted features, making the model’s feature learning more targeted and effective. Combined with the dynamic weighted combined loss that optimizes the training process, the three core modules jointly enhance the model’s segmentation accuracy, boundary consistency and generalization ability. The similar amplitude of performance decline after the removal of each module also proves the rationality of the proposed model architecture design, and that each core module is indispensable for the optimal performance of the model. The superior performance of the full model over other comparison models is the result of the joint action of the independent contribution of each core module and their synergistic effect, which makes the proposed model a strong candidate for clinical spinal image segmentation and classification tasks with higher precision and stability. Table 6. Quantitative results of ablation experiments on lumbar spine MRI dataset and VerSe dataset (%, mean ± standard deviation). Model variant Lumbar spine MRI dataset VerSe dataset Acc P R F1 Acc P R F1 Ours w/o U-ResNet w/o Shape-aware att w/o Dynamic loss Open in a new tab Figure 4 shows how different loss functions affect Discrete Segmentation (DSC) of spine images on two datasets. SPIDER dataset dynamically weighted combined loss works best. Vertebrae DSC intervertebral discs DSC average DSC of both it does much better than other loss functions cross-entropy loss Dice loss included. This shows dynamically weighted combined loss balances vertebrae and intervertebral discs segmentation needs well boosts overall segmentation accuracy. VerSe dataset dynamically weighted combined loss advantages are just as clear. Small vertebrae normal vertebrae abnormal vertebrae it still gets highest DSC beats other functions cross-entropy loss Dice loss among them. This proves its strength in segmenting multiple vertebrae types especially hard targets like small and abnormal vertebrae. It catches feature differences between different vertebrae shapes accurately leads to more reliable segmentation accuracy and completeness. Two datasets results together show spine image segmentation task dynamically weighted combined loss adjusts weights based on actual conditions takes full account of different regions segmentation features. Processes local lumbar spine structures handles all kinds of vertebrae across the spine it works effectively. Finally shows obvious advantage in DSC metric gives strong support for improving model segmentation performance. Fig. 4. Open in a new tab Comparison of DSC scores for different loss functions on the lumbar spine MRI and VerSe datasets. Visual validation of segmentation performance across datasets and modalities From the detailed comparison of the segmentation results in Fig. 5 , it can be clearly seen that there is a significant gap between the segmentation results of the traditional U-Net model and the manual annotation standards of experts. The traditional U-Net model is not precise enough in delineating the boundaries of the vertebral body, and may make errors in identifying certain areas, making it impossible to fully match the anatomical morphology of the lumbar spine. The segmentation results after optimizing the image size show significant improvements in both the fit of the vertebral body boundary and the integrity of the segmentation area. Whether it is the contour of the L2 segment vertebral body or the distinction between the L5 segment and surrounding structures, the optimized results are highly aligned with expert annotations, with almost no boundary deviations or regional omissions. This difference fully demonstrates that image size optimization strategies can effectively address the performance shortcomings of traditional models in lumbar spine segmentation. After adjusting the size of the input image, the model can more accurately capture the morphological features of multiple segments of the lumbar spine, thereby improving segmentation accuracy. This also provides more reliable technical support for the imaging evaluation of lumbar spine diseases in the future. Fig. 5. Open in a new tab Comparison of segmentation effects of different methods on multiple lumbar segments (L1, L2, L5). As can be seen from the Fig. 6 , in the Lumbar Spine MRI dataset section, the model can accurately segment the lumbar intervertebral disc region. The green and red annotations correspond to intervertebral discs in different states, and their contours are highly consistent with the intervertebral disc structures in the MRI grayscale images. This reflects the model’s ability to capture details of local soft tissues in the lumbar spine and clearly distinguish the morphological differences between normal and pathological intervertebral discs. In the VerSe dataset section, the vertebral bodies and intervertebral disc structures annotated with multiple colors have distinct layers, where different colors correspond to different vertebral body or intervertebral disc segments. The model can completely present the multi-segment structure of the entire spine. Whether it is normal vertebral bodies or complex segment combinations, the segmentation boundaries are clear and accurate. The model has good adaptability to the segmentation of complex structures of the entire spine and can effectively cope with morphological changes in different segments such as cervical, thoracic, and lumbar vertebrae. This chart intuitively verifies the effectiveness of the model in different spinal datasets and segmentation tasks of different spinal structures. It can accurately segment the details of local lumbar intervertebral discs and fully present the multi segmental structure of the entire spine, providing visual evidence for the high-precision performance of the model in spinal image segmentation. At the same time, it also shows that the model has good adaptability to two different modalities of spinal images (MRI and CT), which can meet the clinical needs for accurate segmentation of multi-modal spinal images. Fig. 6. Open in a new tab Segmentation visualization comparison of different spine datasets. From Fig. 7 , it can be seen that in the Lumbar Spine MRI dataset, for severe lumbar disc herniation, the model can accurately segment the protruded intervertebral disc tissue and the compressed and deformed spinal canal boundary, with the segmentation contour highly consistent with the anatomical structure. For severe vertebral compression fractures, the model can clearly identify the collapsed and deformed vertebral morphology, and completely retain the boundary details of the fracture area. For severe scoliosis, the model can stably segment the curved vertebral sequence without contour deviation due to anatomical structure displacement. In the VerSe dataset, for the segmentation task of the slender structure of the whole spine, the model can maintain the consistency of the continuous spinal canal and vertebral boundary. For abnormal vertebrae with tumor infiltration, the model can accurately locate the lesion area invaded by tumors, and the segmentation contour does not extend excessively due to blurred boundaries. For small-volume irregular lesions such as vertebral accessory tumors, the model can also completely segment the lesion tissue without missed segmentation or false positive areas. Overall, the model exhibits stable segmentation performance in various severe lesion scenarios, and can effectively cope with complex structural changes caused by diseases in clinical practice, providing a reliable imaging basis for subsequent pathological analysis and clinical decision-making. Fig. 7. Open in a new tab Figure 1 shows the qualitative segmentation results of the model on six typical challenging cases, including Case 1 (severe lumbar disc herniation), Case 2 (severe vertebral compression fracture), Case 3 (severe scoliosis), Case 4 (whole spine multi-structure segmentation), Case 5 (abnormal vertebra with tumor infiltration), and Case 6 (vertebral accessory tumor). Covering the Lumbar Spine MRI and VerSe datasets, the figure includes original images, model segmentation results and 3D reconstruction effects, intuitively demonstrating the model’s segmentation capability for lesions with severe structural changes. Discussion This study addresses key challenges in spinal medical image segmentation, including inadequate handling of complex disease interference, poor adaptability to real clinical settings, and misalignment between segmentation results and clinical requirements. To resolve these issues, we developed an end-to-end deep learning model integrated with a U-ResNet backbone, a shape-aware attention module, and a dynamically weighted combined loss function. The experimental results demonstrate the model’s superior performance: it achieved an accuracy of 87.9% and an F1 score of 85.8 in spinal disease classification, a maximum Dice coefficient of 0.899 for vertebral body segmentation and 0.857 for abnormal vertebrae segmentation, and a top Kappa coefficient of 0.83 in intervertebral disc degeneration grading. These performance metrics outperform those of commonly used comparison models such as U-KAN and CTRCL. Furthermore, our ablation study confirmed the indispensability of the three core components (U-ResNet, shape-aware attention module, and dynamically weighted combined loss function), with the U-ResNet backbone contributing most significantly to the model’s performance improvement. While the proposed model demonstrates promising performance, it is crucial to explicitly acknowledge its limitations, particularly those related to data dependency and potential overfitting, as well as other practical constraints that affect its clinical applicability. First, the model exhibits significant data dependency, as it was trained primarily on the Lumbar Spine MRI and VerSe datasets, which are dominated by common spinal diseases such as lumbar disc herniation, scoliosis, and vertebral compression fractures. These datasets lack sufficient samples of rare spinal diseases and suffer from class imbalance, which directly restricts the model’s generalization ability to diverse clinical scenarios. Specifically, the model may exhibit suboptimal performance when applied to spinal images from different medical institutions with varying scanning protocols, imaging devices, or patient populations, and it struggles to accurately segment and classify rare spinal diseases. Additionally, the limited number of rare disease samples increases the risk of potential overfitting to common disease patterns, which reduces the model’s reliability in real clinical settings where accurate diagnosis of rare cases is critical for patient treatment. Even though we implemented dropout layers and a difficulty-adaptive sampling mechanism to mitigate overfitting, the model still faces inherent overfitting risks when dealing with small-sample pathological subtypes, further compromising its robustness in clinical applications involving heterogeneous patient data and different clinical settings. Second, the model still has room for improvement in terms of computational efficiency and resource consumption. While the current model can effectively complete segmentation and classification tasks, it requires substantial computational resources to operate, which poses challenges for real-time 3D spine segmentation–a key requirement for practical clinical application. Further optimization of the model to reduce memory usage and improve running speed is therefore necessary to facilitate its widespread adoption across various medical institutions. The proposed model holds substantial practical implications for clinical spinal diagnosis, as its superior performance directly addresses unmet needs in clinical practice. It can efficiently and accurately identify pathological spinal regions, reduce the workload associated with manual image analysis, and improve the efficiency and consistency of spinal disease screening and diagnosis. This standardization of diagnostic processes helps narrow the gap in diagnostic quality between different medical institutions and enhances the reliability of clinical decision-making related to spinal diseases. To address the identified limitations and further enhance the clinical value of the proposed method, three concrete and actionable future research directions are outlined. First, we will expand and diversify the training dataset through multi-center collaboration, collecting spinal images with diverse scanning protocols, different imaging devices, and a comprehensive range of pathological types–including rare spinal diseases such as vertebral burst fractures and spinal tumors. This expansion will alleviate data dependency, reduce overfitting risks, and improve the model’s generalization ability to real-world clinical data and diverse clinical settings. Second, we will optimize the model architecture by adopting model compression techniques, lightweight network designs, and quantization strategies to reduce its computational and memory requirements, enabling real-time 3D spine segmentation and promoting its widespread clinical deployment. Third, we will integrate the model more closely with clinical workflows by developing a dedicated clinical decision support system that not only assists in diagnosis but also provides quantitative analysis to support surgical planning and postoperative recovery assessment. We will also conduct prospective clinical validation studies to evaluate the model’s impact on diagnostic accuracy and patient outcomes, refining the model based on feedback from radiologists and spine surgeons to ensure it delivers timely, actionable information that directly improves patient care and treatment results. Additionally, in future work, we will conduct systematic experimental evaluations to quantify the model’s generalization capability across different imaging devices and clinical settings, including multi-center trials with heterogeneous scanning equipment and varied clinical practice standards, to fully verify its practical applicability in real-world clinical environments. Author contributions The authors contributed as follows: Dexuan Zhao contributed to the conceptualization, methodology, software development, and drafting of the manuscript. Rujie Qin, as the corresponding author, was involved in the study design, supervision, and overall project management. Zhijin Chai contributed to data analysis and validation. Shaoxin Ma supported the data curation and visualization, while Qipeng Gao assisted with formal analysis and writing the manuscript. All authors contributed to reviewing and editing the manuscript. Funding Not applicable. Data availability The Lumbar Spine MRI Dataset is available at https://data.mendeley.com/datasets/k57fr854j2/2 and the VerSe dataset is available at https://github.com/anjany/verse . All relevant code and data used in this study are available upon request from the corresponding author. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Degadwala, S., Krishnamurthy, V. N. D. & Vyas, D. Deepspine: multi-class spine x-ray conditions classification using deep learning. In 2024 3rd International Conference on Sentiment Analysis and Deep Learning (ICSADL) , 8–13 (IEEE, 2024). 2. Saeed, M. U., Bin, W., Sheng, J., Albarakati, H. M. & Dastgir, A. Msff: An automated multi-scale feature fusion deep learning model for spine fracture segmentation using MRI. Biomed. Signal Process. Control 91 , 105943 (2024). [ Google Scholar ] 3. Gaikwad, D., Sejal, A., Bagade, S., Ghodekar, N. & Labade, S. Identification of cervical spine fracture using deep learning. Austr. J. Multi-Discip. Eng. 20 , 48–56 (2024). [ Google Scholar ] 4. Hokamura, M. et al. Super-resolution deep learning reconstruction approach for enhanced visualization in lumbar spine MR bone imaging. Eur. J. Radiol. 178 , 111587 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 5. Wan, Y. et al. Machine learning models based on CT radiomics features for distinguishing benign and malignant vertebral compression fractures in patients with malignant tumors. Acta Radiol. 65 , 1359–1367 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 6. Azad, R. et al. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell. (2024). [ DOI ] [ PubMed ] 7. Goceri, E. Polyp segmentation using a hybrid vision transformer and a hybrid loss function. J. Imag. Inf. Med. 37 , 851–863 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Roshan, S. et al. A deep ensemble medical image segmentation with novel sampling method and loss function. Comput. Biol. Med. 172 , 108305 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 9. Ruan, J., Li, J. & Xiang, S. Vm-UNet: Vision mamba unet for medical image segmentation (ACM Transactions on Multimedia Computing, Communications and Applications, 2024). [ Google Scholar ] 10. Wang, Z., Zheng, J.-Q., Zhang, Y., Cui, G. & Li, L. MAMBA-UNet: Unet-like pure visual mamba for medical image segmentation. arXiv preprint arXiv:2402.05079 (2024). 11. Pham, T.-H., Li, X. & Nguyen, K.-D. seUNet-Trans: A simple yet effective UNet-transformer model for medical image segmentation. IEEE Access (2024). 12. Dan, Y., Jin, W., Yue, X. & Wang, Z. Enhancing medical image segmentation with a multi-transformer U-Net. PeerJ 12 , e17005 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Bakkouri, I. & Afdel, K. Multi-scale cnn based on region proposals for efficient breast abnormality recognition. Multimed. Tools Appl. 78 , 12939–12960 (2019). [ Google Scholar ] 14. Bakkouri, I. & Afdel, K. Computer-aided diagnosis (CAD) system based on multi-layer feature fusion network for skin lesion recognition in dermoscopy images. Multimedia Tools and Applications 79 , 20483–20518 (2020). [ Google Scholar ] 15. Zhu, Z. et al. Lightweight medical image segmentation network with multi-scale feature-guided fusion. Comput. Biol. Med. 182 , 109204 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 16. Wang, C. et al. CFATransuNet: Channel-wise cross fusion attention and transformer for 2D medical image segmentation. Comput. Biol. Med. 168 , 107803 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Huang, S., Wang, L., Liao, J. & Liu, L. Multi-attentional causal intervention networks for medical image diagnosis. Knowl.-Based Syst. 299 , 111993 (2024). [ Google Scholar ] 18. Ma, C. & Wang, Z. SEMI-MAMBA-Unet: Pixel-level contrastive and cross-supervised visual mamba-based unet for semi-supervised medical image segmentation. Knowl.-Based Syst. 300 , 112203 (2024). [ Google Scholar ] 19. Rahman, M. M. & Marculescu, R. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Medical Imaging with Deep Learning , 1526–1544 (PMLR, 2024). 20. Ghamsarian, N., Wolf, S., Zinkernagel, M., Schoeffmann, K. & Sznitman, R. DeepPyramid+: medical image segmentation using pyramid view fusion and deformable pyramid reception. Int. J. Comput. Assist. Radiol. Surg. 19 , 851–859 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Liu, T., Bai, Q., Torigian, D. A., Tong, Y. & Udupa, J. K. VSmTrans: A hybrid paradigm integrating self-attention and convolution for 3D medical image segmentation. Med. Image Anal. 98 , 103295 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Zhu, S. et al. SECA-Net: Squeezed-and-excitated contextual attention network for medical image segmentation. Biomed. Signal Process. Control 97 , 106704 (2024). [ Google Scholar ] 23. Wang, L., Xu, P., Cao, X., Nappi, M. & Wan, S. Label-aware attention network with multi-scale boosting for medical image segmentation. Expert Syst. Appl. 255 , 124698 (2024). [ Google Scholar ] 24. Chen, T. et al. xLSTM-UNet can be an effective 2D & 3D medical image segmentation backbone with Vision-LSTM (ViL) better than its mamba counterpart. arXiv preprint arXiv:2407.01530 (2024). 25. Tian, Q., Wang, Z. & Cui, X. Improved unet brain tumor image segmentation based on GSCONV module and ECA attention mechanism. arXiv preprint arXiv:2409.13626 (2024). 26. Zheng, F. et al. SMAFormer: Synergistic multi-attention transformer for medical image segmentation. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , 4048–4053 (IEEE, 2024). 27. Song, E., Zhan, B. & Liu, H. Combining external-latent attention for medical image segmentation. Neural Netw. 170 , 468–477 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 28. Wang, Z., Chen, Y., Wang, F. & Bao, Q. Improved unet model for brain tumor image segmentation based on ASPP-coordinate attention mechanism. In 2024 5th International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE) , 393–397 (IEEE, 2024). 29. Li, X. et al. DMSA-UNet: Dual multi-scale attention makes unet more strong for medical image segmentation. Knowl.-Based Syst. 299 , 112050 (2024). [ Google Scholar ] 30. Nguyen-Tat, T. B., Nguyen, T.-Q.T., Nguyen, H.-N. & Ngo, V. M. Enhancing brain tumor segmentation in MRI images: A hybrid approach using UNet, attention mechanisms, and transformers. Egypti. Inf. J. 27 , 100528 (2024). [ Google Scholar ] 31. Zhong, J. et al. PMFSNET: Polarized multi-scale feature self-attention network for lightweight medical image segmentation. Comput. Methods Programs Biomed. 261 , 108611 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 32. Bakkouri, I., Afdel, K., Benois-Pineau, J. & Initiative, G. C. F. t. A. D. N. BG-3DM2F: Bidirectional gated 3d multi-scale feature fusion for alzheimer’s disease diagnosis. Multimed. Tools Appl. 81 , 10743–10776 (2022). 33. Huang, H. et al. Channel prior convolutional attention for medical image segmentation. Comput. Biol. Med. 178 , 108784 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 34. Zhang, J., Luan, Z., Ni, L., Qi, L. & Gong, X. Msdanet: A multi-scale dilation attention network for medical image segmentation. Biomed. Signal Process. Control 90 , 105889 (2024). [ Google Scholar ] 35. Bakkouri, I. & Afdel, K. MLCA2F: Multi-level context attentional feature fusion for covid-19 lesion segmentation from ct scans. SIViP 17 , 1181–1188 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Zhao, Y., Li, J., Ren, L. & Chen, Z. Dtan: Diffusion-based text attention network for medical image segmentation. Comput. Biol. Med. 168 , 107728 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 37. Gangrade, S., Sharma, P. C., Sharma, A. K. & Singh, Y. P. Modified DeeplabV3+ with multi-level context attention mechanism for colonoscopy polyp segmentation. Comput. Biol. Med. 170 , 108096 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 38. Zhang, X. et al. Smtf: Sparse transformer with multiscale contextual fusion for medical image segmentation. Biomed. Signal Process. Control 87 , 105458 (2024). [ Google Scholar ] 39. Vitale, J., Sconfienza, L. M. & Galbusera, F. Cross-sectional area and fat infiltration of the lumbar spine muscles in patients with back disorders: A deep learning-based big data analysis. Eur. Spine J. 33 , 1–10 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 40. Liu, H., Huang, J., Li, Q., Guan, X. & Tseng, M. A deep convolutional neural network for the automatic segmentation of glioblastoma brain tumor: Joint spatial pyramid module and attention mechanism network. Artif. Intell. Med. 148 , 102776 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 41. Chen, Y. et al. Icycle-gan: Improved cycle generative adversarial networks for liver medical image generation. Biomed. Signal Process. Control 92 , 106100 (2024). [ Google Scholar ] 42. Xu, J. Hc-mamba: Vision mamba with hybrid convolutional techniques for medical image segmentation. arXiv preprint arXiv:2405.05007 (2024). 43. Li, C. et al. U-kan makes strong backbone for medical image segmentation and generation. Proc. AAAI Conf. Artif. Intell. 39 , 4652–4660 (2025). [ Google Scholar ] 44. Chen, J. et al. Transunet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med. Image Anal. 97 , 103280 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 45. Agrawal, A., Agrawal, A., Gupta, S. & Bagade, P. Kan-mamba fusionnet: Redefining medical image segmentation with non-linear modeling. arXiv preprint arXiv:2411.11926 (2024). 46. Wu, L. et al. Cnn-transformer rectified collaborative learning for medical image segmentation. IEEE Trans. Circ. Syst. Video Technol. (2024). 47. Zhou, H.-Y. et al. nnformer: Volumetric medical image segmentation via a 3d transformer. IEEE Trans. Image Process. 32 , 4036–4045 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 48. Zhu, J., Hamdi, A., Qi, Y., Jin, Y. & Wu, J. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv preprint arXiv:2408.00874 (2024). 49. Tang, W., Wu, Z., Wang, W., Pan, Y. & Gan, W. VM-UNet++ research on crack image segmentation based on improved VM-UNet. Sci. Rep. 15 , 8938 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Möller, H. et al. Spineps-automatic whole spine segmentation of t2-weighted mr images using a two-phase approach to multi-class semantic and instance segmentation. Eur. Radiol. 35 , 1178–1189 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Saeed, M. U., Bin, W., Sheng, J. & Saleem, S. 3d mfa: An automated 3d multi-feature attention based approach for spine segmentation using a multi-stage network pruning. Comput. Biol. Med. 185 , 109526 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The Lumbar Spine MRI Dataset is available at https://data.mendeley.com/datasets/k57fr854j2/2 and the VerSe dataset is available at https://github.com/anjany/verse . All relevant code and data used in this study are available upon request from the corresponding author. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (4.1 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• State-of-the-art and future developments in dosimetry for SSTR- and PSMA-targeting radioligand therapies.#1016098
• American Association for Thoracic Surgery Quality Gateway: A surgeon case study of its application in adult cardiac surgery for quality assurance.#1016099
• Metabolic Reprogramming of Cancer Stem Cells: Targeting Lipid Flux and Mitochondrial Plasticity to Overcome Therapeutic Resistance.#1016100
• Symptom Bother Is Associated With Medical Care‐Seeking for LUTS in Japanese Women: A Survey of 1800 Women.#1016101