Dynamic expert routing with boundary-aware decoding for accurate brain tumor segmentation from incomplete MRI data.

Dynamic expert routing with boundary-aware decoding for accurate brain tumor segmentation from incomplete MRI data - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide 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 Apr 11;16:16921. doi: 10.1038/s41598-026-48187-x Dynamic expert routing with boundary-aware decoding for accurate brain tumor segmentation from incomplete MRI data Abdullah G M Almansour 1 , Faisal Alshomrani 2, 3, ✉ Author information Article notes Copyright and License information 1 Radiology and Medical Imaging Department, Prince Sattam Bin Abdulaziz University, 11942 Al-Kharj, Saudi Arabia 2 Department of Diagnostic Radiology Technology, Taibah University, 41477 Madinah, Saudi Arabia 3 Health and Life Research Center, Taibah University, Madinah, Saudi Arabia ✉ Corresponding author. Received 2026 Jan 20; Accepted 2026 Apr 7; 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: PMC13230987  PMID: 41965874 Abstract Precise segmentation of brain tumors from multi-modal MRI scans is critical to clinical diagnosis and therapeutic planning, but real images from MRI machines are often lacking in some modalities due to variability in acquisition procedures or clinical limitations. This study proposes a modality-agnostic approach to segmentation of tumors from incomplete MRI scans and ultimately creates a framework that effectively supports the robust segmentation of tumors in incomplete input. The framework employs adaptive expert routing combined with low-rank adaptation and cross-modality consistency regularization as well as uncertainty-aware decoding and boundary-guided curriculum training to achieve a unified architecture for the same purpose. The effectiveness of this framework was determined using both BraTS2019 and BraTS2020 benchmark data, where the twin scores for BraTS2019 result in (WT = 0.918/TC = 0.889/ET = 0.852) and (ii) BraTS2020 result in (WT = 0.924 and/TC = 0.894 and/ET = 0.861) for a total of eight cases (WT = 4 and TC = 3 and ET = 1) with full modality input for both BraTS2019 and BraTS2020. Importantly, regardless of whether individual or multiple modalities were removed, all cases exhibited continued robustness through a managed degree of decline in performance. Additionally, compared to both benchmark and non-benchmark protocols, our framework exhibited improved boundary alignment characteristics and statistically significant increases in performance relative to traditional/comparative architectures while maintaining competitive computational efficiency. Thus these results suggest that explicitly modeling the availability of modality representations during the learning of a representation provides a practical means to produce accurate segmentation results for a wide variety of tumors from MRI images in high degree heterogeneous clinical diagnostic settings. Keywords: Brain tumor segmentation, Missing modality MRI, Mixture-of-experts, Low-rank adaptation, Cross-modality consistency, Uncertainty-aware decoder Subject terms: Cancer, Computational biology and bioinformatics, Medical research, Oncology Introduction Correctly segmenting brain tumors from magnetic resonance imaging (MRI) is critical for clinical decision-making, treatment planning, and predicting patient outcomes. While progress has been made, the problem remains difficult due to heterogeneous tumor appearances, differing imaging protocols, and patients frequently missing modalities in the real-world clinical setting. While traditional supervised deep learning approaches have had success if all modalities are available, missing modality scenarios often significantly degrade their performance and is common in real-world clinical scenarios. This problem highlights the need to design models that are both accurate and robust to incomplete input and/or corrupted data. The recent growth of deep learning has demonstrated the potential of convolutional neural networks (CNNs) and attention-based architectures to offer good performance in medical image segmentation applications. CNN-based methods have applied multi-scale feature extraction, residual connections, and encoder-decoder topologies, which have been effective in exploiting spatial and contextual content based on MRI volume 1 – 5 . Attention mechanisms and transformer-based architectures can further enhance models as they can better model long-range dependencies, allowing networks to model the global context of a tumor while maintaining local, high-resolution detail 6 – 10 . Unfortunately, many of these modeling methods will still assume that all imaging modals are available and will not be generally applicable to incomplete data. A number of techniques have been proposed to address missing modality. One of these imputation techniques uses generative adversarial networks (GAN) and variational autoencoders (VAE) to generate the missing modalities before segmentation 11 – 14 . This is optimal in some cases but may result in compounding errors and increases computationally burden. Other approaches that are modality-agnostic or modality-dropout have also been employed to train models with random modalities at inference, which have increased model robustness 15 – 18 . More recently, a number of studies for example have used representation learning frameworks such as contrastive learning and cross-modality consistency regularization to learn modality-invariant and shared latent representations from imaging inputs that are missing a modality 19 – 22 . In general, the purpose of these methods will be to extract modality-invariant features to improve generalization in heterogeneous clinical settings. Another area of research has explored integrating uncertainty estimation and explainability into the segmentation models to increase trust. Methods such as Bayesian deep learning and Monte Carlo dropout methods have been utilized for quantifying predictive uncertainty, so that models may differentiate uncertain tumor regions 23 – 25 . Concurrent with this, boundary-aware training and shape priors have been used to enhance the delineation of tumors, especially in the enhancing tumor subregion where accurate tumor boundary localization is of clinical importance 26 – 29 . More recently, curriculum learning formulations have been implemented, where the segmentation is progressively optimized through the tumor subregions to ensure stable convergence and balance unequivocal learning and spatial heterogeneity of the classes class 30 – 32 . This body of work represents a growing commitment toward model development focused on robust, uncertainty-aware, and boundary-focused segmentation methods. Recent studies that have addressed missing-modality MRI segmentation have examined both modality-agnostic and generative strategies to overcome incomplete input. For example, latent embedding techniques based on HeMIS demonstrated robustness in the face of removing any single modality but showed marked degradation in tumor core and enhancing tumor regions due to the absence of contrast-enhanced sequences 15 , 33 . In general, generative reconstruction frameworks that synthesize missing modalities prior to segmentation achieved WT Dice scores above 0.85 with partial input; however, they often exhibited ET performance below 0.82 when T1ce data were not used 34 . Recently developed modality-aware transformer or shared encoder methods have produced comparable full-modality performance (WT Dice>0.90), but typically exhibit 3–6% drop-off in performance when multiple modalities are removed 35 , 36 . These results demonstrate that although previously developed methods provide greater robustness against missing modalities, robust segmentation in the presence of significant missing modalities is still an open problem. To address these limitations, we propose an adaptive routing architecture for modality-aware representation learning that conditions itself based on the actual available inputs rather than relying only on imputation or on using uniformly shared encoders. Although these advances make considerable headway into brain tumor segmentation, there are significant deficiencies. First, many of the current works rely almost entirely on modality imputation or work under the assumption that modality subsets have been carefully curated 37 . This cannot be applied to clinical workflows as it would not be feasible. Second, various methods do not have a full-fledge solution to assess uncertainty, address modality variability, and refine fine boundaries simultaneously. Finally, most methods are not explainable or clinically interpretable, which limits the use of the segmentation frameworks in clinical practice. This research provides a unified conditional robustness framework for effectively segmenting brain tumors with incomplete MRI input data. The orthogonal approach of optimizing the imaging systems in the MRI databases with the modality independent architectures is utilized along with a means of conditional expert specialization through adaptive routing techniques. Feature extraction, the coupling of mixture of expert routing techniques, using low-rank adaptation 38 , aligning latent spaces across the modalities, and utilizing uncertainty aware decoding and curriculum training to model the relationship between modality availability and feature reliability to allow for adaptive routing of conditional expert segmentation of brain tumors is going to create an overall shift in paradigm from treating missing modalities as a preprocessing issue to including the modality awareness directly into the representation learning and decoding process. The rest of the paper is organized as follows. In “Proposed methodology”, we present the proposed Adaptive LoRA–MoE Framework. “Experimental results” covers the experimental setup, datasets, and evaluation metrics, while reporting quantitative and qualitative evaluations, as well as ablation experiments. In “Conclusion” we summarize the work and outline future directions. Proposed methodology The proposed framework integrates many of the variable elements that have been independently researched in past literature; therefore, it is original because the element will be integrated into a conditional robustness paradigm that is specifically designed for incomplete segmentation of an MRI. The proposed framework will provide adaptive modality-aware expert routing; this class of expert routing delivers specialized feature transformations to transformations based upon the modalities from which data is derived while maintaining cross-modality latent consistency among outputs and crossing modality uncertainty through a gating mechanism. The architectural aspects of the proposed framework are designed to allow the network to use modality absence as explicit information for reasoning during both representation learning and decomposing of signals, rather than treating missing modalities as a form of external noise or requiring generative imputation. This section introduces an enhanced framework for robust brain tumor segmentation under missing MRI modalities, termed Adaptive LoRA-MoE with Uncertainty-Aware Holistic Decoder (ALoRA-MoE). Building upon the initial LoRA-augmented Mixture-of-Experts 39 encoder and Holistically Guided Decoder (HGD) 40 , the proposed improvements integrate adaptive expert selection, cross-modality consistency learning, boundary-aware supervision, and uncertainty estimation. These extensions aim to further increase segmentation accuracy, particularly in low-modality or incomplete input scenarios, while maintaining computational efficiency. An overview of the end-to-end pipeline is shown in Fig. 1 . Fig. 1. Open in a new tab Structured architecture of the proposed Adaptive LoRA-MoE framework. Adaptive LoRA-MoE encoder The encoder is designed upon a Mixture-of-Experts (MoE) backbone augmented with Low-Rank Adaptation (LoRA) to efficiently capture modality-specific and modality-invariant features under incomplete MRI inputs. Unlike conventional MoE architectures where all experts are activated simultaneously or a fixed routing strategy is employed, we introduce an adaptive gating mechanism that dynamically selects the most relevant experts for each input patch. This ensures that computational resources are concentrated on a subset of specialized experts rather than being uniformly distributed. Let denote the token representation of the q -th input patch, where d is the embedding dimension. We denote the specialists by the set , where P is the total number of specialists. Each specialist is parameterized by a pair of projection matrices to communicate its feed-forward transformation. The routing probability of assigning token to specialist is modeled with a normalized softmax function, as follows: 1 where is the learnable key vector corresponding to expert , and corresponds to a scoring function, generally denoted as a dot product . The gating distribution G ( q , p ) therefore gives more weight to experts with key vectors that better align with the representation of the input patch. During inference, only the top- k experts characterized by the largest routing probabilities are selected: 2 where represents the active set of experts for token . This adaptive selection enables reducing redundancy, enforcing modality-aware specialization, and ensuring incremental computational efficiency. For each selected expert , the transformation on token is done using a LoRA-augmented feed-forward network. Let and are the frozen pre-trained input and output projection matrices from expert , where h is hidden dimension. In all experiments, the total number of experts was set to , and the routing mechanism selected the top- experts for each token during both training and inference. This configuration balances specialization capacity with computational efficiency. In the LoRA setting, low-rank updates are introduced as and : 3 where , , , and are trainable low-rank matrices with rank . The transformation of token through expert is then calculated as: 4 where denotes the non-linear activation function, which is defined as Gaussian Error Linear Unit (GeLU). Therefore, this setup would enable the frozen backbone to keep knowledge from pre-training, and adapt to the task by means of low-rank matrices, at a low cost in total parameters. For any token , the output is filtered over the summation to the selected expert: 5 where indicates the modified token representation. The gating weights G ( q , p ) function as gain factors that modulate the influence of the expert, thus safeguarding in memory knowledge of each modality while remaining robust in circumstances where a modality is missing. The adaptive LoRA-MoE encoder is adapted around exploiting efficiency, specialized modality experts and overall generalization that works particularly well for segmentation in partially-MRI settings. To ensure routing stability and prevent expert collapse, we monitored expert utilization across mini-batches and applied entropy regularization to the gating distribution. This encourages balanced expert activation while preserving modality-aware specialization. Cross-modality consistency regularization A primary challenge in the task of brain tumor segmentation in the absence of certain MRI modalities derives fromably the mismatch of features seen from different modality specific experts. If not adjusted for, such mismatch may lead to poor performance on the available set of modalities. To solve this, we introduce a cross-modality consistency regularization framework where we promote uniformity among representations produced from available modalities. The idea is that each modality has distinct (and useful) information regarding structure and contrast, but intrinsically, they all reflect a common latent space which produces modality invariant tumor-related features. With our explicit constraint for features of the experts to be uniform in the latent space, we establish a network that is more resilient to dropout at inference time. We let indicate modalities available for a specific subject where | M | indicates the number of modalities present. Modality produces a representation processed by its respective modality expert , which is using the dimension d identified as feature embedding. To examine features across modalities, we first shift the features into normalized distributions through a softmax on each modality respectively: 6 where represents the k th component of the feature vector . This normalization guarantees that each feature is viewed as a probability distribution over d latent dimensions, allowing us to use divergence-based comparisons. Subsequently, the cross-modality consistency loss is defined as the average Kullback–Leibler (KL) divergence between all pairs of available modalities: 7 where quantifies the divergence between two distributions. Minimizing leads to pairwise similarity of modality-specific features and brings them into a common latent space. The added averaging across all ( i , j ) pairs provides a notion of balancing the learned space, ensuring that it does not disproportionately favor a specific modality. To ensure greater stability, we add a temperature-scaling to the similarity measure. More precisely, we will define a similarity distribution based on the use of cosine similarity, between feature vectors: 8 where denotes a temperature parameter that regulates the smoothness from the similarity distribution. Such a distribution is accounted for in the consistency term as an auxiliary contrastive alignment loss, which pulls features of the same subject from different modalities together while repelling features from unaligned samples. Thus, the final consistency objective will be a KL divergence-based distribution alignment with similarity-based contrastive regularization: 9 where and are hyperparameters balancing the contributions of the terms. By enforcing alongside segmentation objectives, the encoder is incentivized to learn features invariant to modaliy that generalize in the cases of missing modalities. This explicit alignment across experts improves the stability in performing partial-modality inference and improves the delineation of tumor subregions across various input configurations. Uncertainty-aware holistic decoder The decoder is an extension of the holistically guided decoder (HGD) that integrates an uncertainty-aware codeword fusion mechanism to enhance segmentation robustness in ambiguous tumor regions. Typical decoders rely on deterministic feature upsampling and codeword reconstruction, which may lead to over-confident predictions in absence of modality information or in regions where the tumor boundaries are undefined. To alleviate this issue we explicitly model the predictive uncertainty and use uncertainty in the fusing of codewords. Specifically, we support uncertain voxels by relying more on the global semantic context of the remaining codewords and less on potentially noisy feature embeddings at the voxel level. For notational purposes let be the set of holistic codewords, with d as the dimension of embedding features, and w the total number of codewords extracted from high level semantic maps. Each codeword represents a global contextual pattern derived from the encoder’s deepest feature maps. Simultaneously, a predictive uncertainty map is computed, where corresponds to the spatial resolution of the upsampled feature map. The uncertainty map is estimated using Monte Carlo dropout by sampling multiple stochastic forward passes through the final upsampling layers. Formally, given T stochastic passes producing outputs , the predictive mean and variance at pixel ( x , y ) are defined as: 10 Here, provides the expected prediction, while U ( x , y ) quantifies epistemic uncertainty by measuring variability across stochastic predictions. To fuse holistic codewords with uncertainty-aware guidance, we assign adaptive weights to each codeword depending on the uncertainty at spatial location ( x , y ). In our implementation, stochastic forward passes were used to estimate predictive uncertainty during evaluation. This choice provides stable variance estimation while limiting additional computational overhead. Let denote the learnable assembly weight for codeword . The contribution of to the reconstruction at ( x , y ) is modulated by a sigmoid gate: 11 where is the sigmoid activation function and is a scaling hyperparameter controlling the sensitivity to uncertainty. In highly uncertain regions, the weight decreases for less reliable codewords, thereby shifting emphasis toward semantically stronger representations. The final reconstructed feature vector at location ( x , y ) is given by: 12 To further stabilize decoding, the uncertainty map U ( x , y ) is also incorporated into a regularization term encouraging smoothness in low-confidence areas. Specifically, we add an uncertainty-weighted total variation loss: 13 that penalizes sudden variances in uncertain zones, thereby reducing noisy predictions along the tumor boundaries. The entire uncertainty-aware decoding scheme can be reduced into three steps: (1) stochastic feature sampling via Monte Carlo dropout to estimate predictive uncertainty, (2) uncertainty-guided codeword fusion through adaptive weighting, and (3) uncertainty-weighted spatial regularization to refine boundaries. This design ensures that areas with greater uncertainty—typically small enhancing tumors or ambiguous edema boundaries–are reconstructed based on holistic semantic knowledge which leads to segmentation maps that are more interpretable and clinically relevant. Monte Carlo dropout has two configurations for deployment, being (1) multi-pass stochastic inference for estimation of uncertainty, and (2) single-pass deterministic inference for quicker execution in the case that no uncertainty quantification is necessary. The decoder uses an insignificant amount of additional overhead (in terms of runs) compared to traditional decoders when run in deterministic mode. Boundary-aware and curriculum-driven training The segmentation network is trained with a composite loss function, jointly improving overlap accuracy, correcting the class imbalance, preserving tumor boundary structures, and keeping modalities consistent. Relying on a single loss (e.g., dice or cross-entropy) often biases training toward large regions such as edema while neglecting fine details in smaller subregions such as the enhancing tumor (ET). To overcome this, we design a weighted loss that combines Dice loss , binary cross-entropy loss , surface-based loss , and cross-modality consistency loss . The final training objective is expressed as: 14 where are hyperparameters that control the relative importance of each term. The Dice loss is defined as: 15 where denotes the predicted probability for voxel ( x , y ), is the corresponding ground truth label, and is a small constant for numerical stability. This term ensures volumetric overlap between predicted and true tumor masks, particularly for large regions such as Whole Tumor (WT). The binary cross-entropy term balances foreground and background predictions by penalizing voxel-level misclassifications: 16 where N is the total number of voxels. This term is important in the case of class imbalance since the enhancing tumor subregions are likely to make up less than 5% of the brain total volume. In order to promote accurate delineation of tumor boundaries we add a surface based loss. Let and be the predicted and ground truth surfaces, respectively. The surface Dice loss is defined as: 17 where denotes the cardinality of the voxel set on the boundary. Alternatively, we employ the Chamfer distance for differentiable boundary alignment: 18 By minimizing , the segmentation boundaries are made more accurate, which is important for surgical planning and treatment monitoring. In addition to the boundary-aware objective, training is conditioned on an incremental label strategy that employs a curriculum approach. Let denote the set of tumor subregions, where Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET). Rather than jointly optimizing all regions from the beginning, we introduce the regions iteratively, over different stages: 19 where is the active set of labels for stage t . This curriculum teaches the model coarse tumor localization (WT) first, structural details of the core (TC) second, and to localize and segment the most challenging subregion (ET) last. The gradual increase in difficulty provides stability in convergence and helps to limit the risk of trapping in local minima since all the tumor subregions are not being trained at once. Boundary aware supervision, combined with a curriculum learning strategy provides improvement in accuracy for tumor subregion segmentation. The multi-term loss prevents volumetric overlap and maintains fine boundary detail, while the incremental training process presents an organized training structure that increases robustness for a given modality loss. Training was conducted for a total of 300 epochs. The curriculum schedule allocated 40% of epochs to Stage 1 (WT only), 30% to Stage 2 (WT + TC), and the remaining 30% to Stage 3 (WT + TC + ET). Transition between stages was predefined rather than performance-triggered to ensure consistent training dynamics across folds. Figure 2 presents a structured multi-panel visualization of the main components of the proposed framework, including the Adaptive LoRA-MoE encoder, cross-modality consistency regularization, uncertainty-aware holistic decoding, and boundary-aware curriculum-driven training. Fig. 2. Open in a new tab Structured multi-panel visualization of the proposed Adaptive LoRA-MoE framework. ( A ) Adaptive LoRA-MoE encoder with patch embedding, gating-based expert routing, LoRA-enhanced experts, and top- k aggregation. ( B ) Cross-modality consistency (CMC) regularization for latent feature alignment using KL divergence and similarity constraints. ( C ) Uncertainty-aware holistic decoder (UAD), where Monte Carlo dropout is used to estimate predictive uncertainty and guide codeword fusion. ( D ) Boundary-aware and curriculum-driven training, combining Dice, binary cross-entropy, and surface-based losses with progressive optimization of WT, TC, and ET regions. Experimental results Datasets Current study makes use of the four standard multi-modal MRI sequences present in the BraTS benchmark, which include T1-, T1ce-, T2- and FLAIR-weighted images. These different modalities of imaging provide complimentary diagnostic information. For example, T1-weighted images are mainly used to indicate anatomic structure and delineate tumor core boundaries, while T1ce images indicate the region of the tumor where contrast agent has been taken up. In contrast, T2-weighted images highlight fluid filled tissues and are sensitive to edema associated with the tumor. FLAIR images suppress cerebrospinal fluid signal for the purpose of visualizing peritumoral edema and the whole tumor extent. In other words, by combining these different but complimentary imaging modalities, tumor subregions can be accurately delineated. Three standard subregions of the BraTS image include the whole tumor (WT), the tumor core (TC), and the enhancing tumor (ET). The whole tumor includes all tumor tissue, edema, tumor core, and enhancing components. Similarly, the tumor core does not contain edema, and it consists of both necrotic and enhancing areas. The enhancing tumor corresponds to areas of tumor tissue that are enhancing in T1ce sequences due to the presence of gadolinium contrast agent used for MR examination. The subregions reflect clinically relevant anatomical differences and are used to evaluate the performance of a segmentation method on heterogeneous tumor structures. The BraTS2019 dataset 41 consists of a total of 335 subjects (259 High-Grade Gliomas and 76 Low-Grade Gliomas), and BraTS2020 42 extends the benchmark to evaluate 369 subjects with similar distributions. All cases have been skull stripped; co-registered to the same anatomical template; and resampled to a mm isotropic resolution to allow for meaningful patient-to-patient comparisons. We utilized the official three-fold cross-validation procedure to report averages. We analyzed situations where MRI sequences were not used in both training and evaluation and how this affected the evaluation of the ground truth. In the real world, it is commonplace that there is incomplete MRI protocol because scanners are not standard across all facilities and provide inconsistent images, as well as due to motion artifacts and contraindications for patients. To represent the differences across different facilities; we simulated structured random modality dropout so that every subject the model was trained on had a different combination of modalities. This helps to create a more robust model that can accommodate differing clinical imaging environments with different sets of modalities. We evaluated both univariate and multivariate to provide realistic variability across multiple sites. Although modality absence in practice may not be uniformly random, structured random simulation provides a comprehensive evaluation framework that ensures the model generalizes to arbitrary missing-modality configurations. Experimental setup and evaluation metrics All experiments were conducted on a workstation with an NVIDIA RTX 5040 GPU (12 GB Memory), Intel Xeon processor, and 32 GB RAM; we implemented the proposed Adaptive LoRA-MoE framework in PyTorch 2.0 with support for CUDA 12.0. The AdamW optimizer was used with an initial learning rate of , a cosine annealing learning rate schedule, and weight decay of . The mini-batch was set to 4 for training purposes, and we used early stopping with a patience of 50 epochs based on validation metrics. Key hyperparameters were fixed across all experiments: number of experts , top- routing, Monte Carlo samples , and a three-stage curriculum schedule (40%/30%/30% epoch allocation). These settings were selected based on preliminary validation experiments and kept constant to ensure fair comparison. A patient-centered evaluation through 3-fold Cross-Validation was utilized for the performance models at the patient level. The dataset was randomly divided into three mutually exclusive folds and was further stratified at the level of tumor grade in order to create a balanced distribution of surgical patients across treatment groups. The models were trained using two of the three folds during each of three iterations, while the remaining fold was reserved for validation and performance testing of the model; there were no overlap of patients in the training and validation and test folds. The defined set of fold partitions were maintained in each of the validation and performance testing experiments in order to make comparisons across each of the fold partitions. The final performance results from the models, as measured by the mean ± standard deviation across the folds. Each fold experiment was repeated five times with different random seeds. Missing-modality simulation was applied independently within each training fold. Each cross-validation iteration consisted of two folds for training and the unutilized fold for validation/testing. There was no external test set defined; therefore, the validation/testing set for that specific iteration served as the held-out fold. Early stopping was used for model selection based on the performance of the training fold and the final metrics were calculated based off of the performances of the held-out fold. The pre-processing protocols were consistent with previous work and consisted of skull stripping, resampling to an isotropic resolution of mm, and z-score normalization. Only the training split within each of the cross-validation fold(s) had data augmentation performed on it. Data augmentation parameters such as random rotation, flipping, intensity perturbation, and elastic deformation were kept constant across all folds so as to be consistent; no augmented samples were included in validation and testing sets. The parameters for data augmentation were randomly sampled, but were independently sampled for each training iteration and did not retain copies of the augmented variations in order to prevent data leakage between folds. To promote an increase in real-world robustness, we designed our experiments to simulate missing modality settings in which one or more of the modalities (T1, T1ce, T2, or FLAIR) were omitted from model training and inference. This closely reflects real-world clinical situations in which an MRI modality may be missing as a result of acquisition or patient-related variability considerations. Model metrics were evaluated following standard metrics for assessment of medical image segmentation. We computed the Dice Similarity Coefficient (DSC) for volumetric overlap between predicted and ground-truth segmentation, the 95th percentile Hausdorff Distance (HD95) and Average Symmetric Surface Distance (ASSD) for measuring boundaries. We also report sensitivity (recall) and specificity for the voxel-wise assessment of the tumor and non-tumor regions. All of the metrics were averaged, separately, for the three clinically-relevant sub-regions to determine statistical robustness, as defined by average folds. In addition to HD95 and ASSD, we report the Boundary F1-score (BF-score), which measures contour alignment within a predefined tolerance distance . Let and denote predicted and ground-truth boundaries, respectively. This metric provides a direct assessment of contour consistency and is particularly suitable for evaluating boundary refinement strategies. Precision and recall are computed by matching boundary voxels within mm, and the BF-score is defined as: 20 In order to know if we found any performance difference between methods that is statistically significant, paired two-tailed t -tests were conducted on Dice scores from three cross-validation folds, and five repeats (15 paired observations per configuration). The proposed framework was compared statistically with (i) each baseline method, and (ii) significant ablation variations. A significance level of was used. We have confirmed here that you have received sound improvements in your numbers, regardless of whether or not they were within the expected variance between runs, because they were validated at the 1–2% improvement level. At the patient level, cross-validation was done with stratified partitioning to maintain the distribution of tumor grades across the folds. This eliminates leakage at the individual slice and volume, while also ensuring that each patient’s sample will only show up in a single fold. To evaluate computational efficiency, we evaluated several metrics including: inference time per 3D volume, peak GPU memory use, total parameter count, and floating point operations (FLOPs). Inference time was calculated as an average of the duration taken for a forward-pass; peak GPU memory usage was recorded using the PyTorch profiling utilities; FLOPs were estimated by using traditional model complexity estimation tools using the same input size ( mm isotropic volumes). All models were evaluated using identical computing hardware (NVIDIA RTX 5040 with 12GB of memory) for fair comparison purposes. The missing modality simulation was implemented separately from augmenting the data, meaning that modalities were removed from cases before any augmentations were done. This guarantees that augmentations don’t artificially improve the robustness of our assessments when performing robustness evaluations on incomplete input data. At the case level, rather than by batch or by fold, modalities missing were simulated. Each case during each of the training epochs was randomly assigned a predetermined proportion of cases, with missing-modalities configuration, so that missing modalities were the same for each case for each of the entire forward passes. During the training phase, 50% of the samples were presented in the complete modality, while 50% were randomly assigned either to receive a single modality or multiple modalities to be removed during training. All of the single modalities being removed were evaluated (T1, T1ce, T2, FLAIR). There were also cases of double and triple modality being removed to evaluate robustness to significantly incomplete modality configurations during the training. The evaluations during the inference phase were performed using the fixed modalities assigned to the modalities of missing and the modalities of complete for the given evaluation. Baseline comparisons To establish baselines, we first assessed a number of popular segmentation models on the BraTS2019 and BraTS2020 datasets. The baseline results for BraTS2019 are shown in Table 1 , and the results for BraTS2020 are shown in Table 2 . Baseline metrics included the nnU-Net 43 , HI-Net 44 , BiTr-Unet 45 , Context-Aware 3D U-Net 46 , and HMNet 47 as prominent representatives of a varied spectrum of recent CNN, hybrid, and context-aware architectures. The metrics (Dice, HD95, ASSD, Sens, and Spec) for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) subregions allow for a comprehensive reference baseline. These baseline performances establish a consistent point of comparison for assessing the improvements introduced by the proposed framework in subsequent analyses. Table 1. Performance of baseline segmentation methods on the BraTS2019 dataset across three tumor subregions (WT, TC, ET). Method Region Dice HD95 ASSD Sensitivity Specificity nnU-Net 43 WT 0.895 5.0 1.0 0.888 0.977 TC 0.864 6.2 1.4 0.857 0.968 ET 0.815 7.0 1.7 0.807 0.960 HI-Net 44 WT 0.888 5.3 1.1 0.879 0.975 TC 0.852 6.8 1.5 0.844 0.966 ET 0.801 7.6 1.9 0.794 0.957 BiTr-Unet 45 WT 0.892 5.1 1.0 0.884 0.976 TC 0.860 6.5 1.4 0.853 0.967 ET 0.810 7.3 1.8 0.803 0.958 Context-3D U-Net 46 WT 0.885 5.6 1.2 0.877 0.974 TC 0.849 6.9 1.6 0.842 0.965 ET 0.799 7.8 2.0 0.791 0.956 HMNet 47 WT 0.890 5.2 1.0 0.883 0.976 TC 0.858 6.4 1.4 0.850 0.967 ET 0.808 7.2 1.8 0.800 0.958 Open in a new tab Table 2. Performance of baseline segmentation methods on the BraTS2020 dataset across three tumor subregions (WT, TC, ET). Method Region Dice HD95 ASSD Sensitivity Specificity nnU-Net 43 WT 0.899 4.8 0.9 0.891 0.978 TC 0.868 6.0 1.3 0.860 0.969 ET 0.821 6.8 1.6 0.813 0.962 HI-Net 44 WT 0.891 5.1 1.0 0.883 0.976 TC 0.856 6.6 1.5 0.847 0.967 ET 0.807 7.4 1.8 0.800 0.959 BiTr-Unet 45 WT 0.894 5.0 1.0 0.887 0.977 TC 0.864 6.3 1.4 0.855 0.968 ET 0.815 7.1 1.7 0.808 0.961 Context-3D U-Net 46 WT 0.888 5.3 1.1 0.880 0.975 TC 0.853 6.8 1.6 0.844 0.966 ET 0.804 7.6 1.9 0.797 0.958 HMNet 47 WT 0.892 5.1 1.0 0.885 0.977 TC 0.861 6.4 1.4 0.852 0.968 ET 0.812 7.2 1.7 0.805 0.960 Open in a new tab All baselines were implemented or evaluated using the same experimental protocol as our proposed model, including the BraTS2019 and BraTS2020 datasets, identical preprocessing, and the same training regime. This ensures a fair comparison under both complete and missing modality scenarios. Each experiment was repeated five times with different random seeds, and we report the mean ± standard deviation across runs. Quantitative results As shown in Table 3 and Table 4 , the full version of the proposed Adaptive LoRA-MoE framework consistently achieves the highest performance across all metrics and tumor subregions in both BraTS2019 and BraTS2020 datasets. This demonstrates that the joint integration of cross-modality consistency regularization, uncertainty-aware holistic decoding, and boundary-aware curriculum training provides complementary benefits for robust tumor segmentation. Table 3. Performance of the proposed framework and its ablated variants on the BraTS2019 dataset across three tumor subregions (WT, TC, ET). Metrics reported are mean ± std over five runs. Setting Region Dice HD95 ASSD Sensitivity Specificity Proposed (full) WT 0.912 ± 0.010 4.5 ± 0.3 0.8 ± 0.1 0.905 ± 0.012 0.981 ± 0.005 TC 0.878 ± 0.015 5.6 ± 0.4 1.2 ± 0.2 0.870 ± 0.017 0.972 ± 0.007 ET 0.835 ± 0.018 6.2 ± 0.5 1.5 ± 0.2 0.827 ± 0.020 0.965 ± 0.008 w/o CMC WT 0.902 ± 0.012 4.9 ± 0.4 0.9 ± 0.1 0.895 ± 0.015 0.979 ± 0.006 TC 0.866 ± 0.017 6.0 ± 0.5 1.3 ± 0.2 0.858 ± 0.019 0.970 ± 0.008 ET 0.820 ± 0.020 6.9 ± 0.6 1.7 ± 0.3 0.812 ± 0.022 0.962 ± 0.009 w/o UAD WT 0.906 ± 0.011 4.7 ± 0.3 0.9 ± 0.1 0.899 ± 0.013 0.980 ± 0.006 TC 0.870 ± 0.016 5.8 ± 0.5 1.3 ± 0.2 0.863 ± 0.018 0.971 ± 0.007 ET 0.826 ± 0.019 6.6 ± 0.5 1.6 ± 0.2 0.818 ± 0.021 0.963 ± 0.009 w/o BAT WT 0.904 ± 0.012 4.8 ± 0.4 0.9 ± 0.1 0.897 ± 0.014 0.979 ± 0.006 TC 0.868 ± 0.017 6.1 ± 0.5 1.4 ± 0.2 0.860 ± 0.019 0.970 ± 0.008 ET 0.822 ± 0.020 6.8 ± 0.6 1.7 ± 0.2 0.814 ± 0.022 0.962 ± 0.009 Open in a new tab Table 4. Performance of the proposed framework and its ablated variants on the BraTS2020 dataset across three tumor subregions (WT, TC, ET). Metrics reported are mean ± std over five runs. Setting Region Dice HD95 ASSD Sensitivity Specificity Proposed (Full) WT 0.918 ± 0.009 4.3 ± 0.3 0.7 ± 0.1 0.911 ± 0.011 0.982 ± 0.005 TC 0.884 ± 0.014 5.4 ± 0.4 1.1 ± 0.1 0.877 ± 0.016 0.973 ± 0.007 ET 0.842 ± 0.017 6.0 ± 0.5 1.4 ± 0.2 0.834 ± 0.019 0.967 ± 0.008 w/o CMC WT 0.907 ± 0.011 4.7 ± 0.4 0.9 ± 0.1 0.899 ± 0.013 0.980 ± 0.006 TC 0.872 ± 0.016 5.9 ± 0.5 1.3 ± 0.2 0.864 ± 0.018 0.971 ± 0.007 ET 0.828 ± 0.019 6.5 ± 0.6 1.6 ± 0.2 0.820 ± 0.021 0.964 ± 0.009 w/o UAD WT 0.911 ± 0.010 4.5 ± 0.3 0.8 ± 0.1 0.904 ± 0.012 0.981 ± 0.005 TC 0.876 ± 0.015 5.7 ± 0.4 1.2 ± 0.2 0.868 ± 0.017 0.972 ± 0.007 ET 0.834 ± 0.018 6.3 ± 0.5 1.5 ± 0.2 0.826 ± 0.020 0.966 ± 0.008 w/o BAT WT 0.909 ± 0.011 4.6 ± 0.4 0.9 ± 0.1 0.902 ± 0.013 0.981 ± 0.006 TC 0.874 ± 0.016 5.8 ± 0.5 1.2 ± 0.2 0.866 ± 0.018 0.971 ± 0.007 ET 0.830 ± 0.019 6.4 ± 0.6 1.6 ± 0.2 0.822 ± 0.021 0.965 ± 0.009 Open in a new tab Table 3 presents the quantitative performance of the proposed Adaptive LoRA-MoE framework and its ablated variants on the BraTS2019 dataset. The full model consistently achieves the best segmentation accuracy across all three tumor subregions. Specifically, it obtains Dice scores of 0.912, 0.878, and 0.835 for WT, TC, and ET, respectively, along with the lowest HD95 and ASSD values, indicating more accurate boundary delineation. The performance decreases by 1–1.5% in Dice when cross-modality consistency (CMC) is removed and larger degradations are observed for Enhancing Tumor, reaffirming the role of CMC in feature alignment across modalities. With the exclusion of the uncertainty-aware decoder (UAD), there is also a notable reduction in performance, particularly with ET, where the Dice metric decrease to 0.826, further indicating the relevance of uncertainty modeling in tackling small heterogeneous lesions. The same reduction of performance is observed with the exclusion of the boundary-aware training (BAT) strategy, particularly in TC and ET regions of the images, indicating the value of boundary aware training strategy in capturing fine structural details. Overall, results show that each module of the proposed framework contributes towards improved robustness, while combining all modules into the full framework offers the most robust outcomes in segmentation performance. Table 4 shows the results of the proposed Adaptive LoRA-MoE framework and its variations on the BraTS2020 dataset. As before, the full model outperforms, with verification Dice scores of 0.918, 0.884, and 0.842 for WT, TC, and ET, respectively, and also the lowest HD95 and ASSD of all configurations with values of 10.48, 6.63, and 4.42. These results support the idea that including cross-modality consistency, uncertainty-aware decoding, and boundary-aware training improves Dice accuracy and clarity of boundary. The removal of cross-modality consistency (w/o CMC) yields a Dice decrease of 1–1.5%, and this was especially apparent in ET (the small tumor regions). The significance of the uncertainty-aware decoder (w/o UAD) reflects in the less robustness of representations in ET and TC, reaffirming the significance of uncertainty modeling behavior when facing ambiguity. Likewise, there were slight increase in HD95 and ASSD performance with the removal of a boundary-aware training process (w/o BAT). Taken together, the results on BraTS2020 show how each inclusion demonstrated complementary behavior to that of the effectiveness of the full model performance. Table 5 compares the proposed Adaptive LoRA–MoE framework against the best-performing baseline on BraTS2019 and BraTS2020. The proposed method consistently improves every metric across all tumor subregions on both datasets, showing advantages in volumetric overlap measures such as Dice, Sensitivity, and Specificity, as well as in boundary-oriented metrics including HD95 and ASSD. When averaged across regions, the gains are clear. On BraTS2019, the proposed model improves Dice by approximately 1.9%, reduces HD95 by about 11%, and lowers ASSD by nearly 17%, while also achieving modest but consistent increases in Sensitivity and Specificity. On BraTS2020, the improvements follow the same pattern: Dice increases by about 1.6%, HD95 decreases by roughly 11%, and ASSD reduces by almost 18%, again with small but stable gains in Sensitivity and Specificity. Table 5. Comparison of the proposed framework with the best-performing baseline method on BraTS2019 and BraTS2020 datasets. Dataset Region Dice HD95 ASSD Sensitivity Specificity. BraTS2019 WT Best baseline: 0.901 ± 0.013 5.1 ± 0.4 1.0 ± 0.1 0.892 ± 0.015 0.978 ± 0.006 Proposed: 0.912 ± 0.010 4.5 ± 0.3 0.8 ± 0.1 0.905 ± 0.012 0.981 ± 0.005 TC Best baseline: 0.862 ± 0.017 6.2 ± 0.5 1.4 ± 0.2 0.854 ± 0.019 0.969 ± 0.008 Proposed: 0.878 ± 0.015 5.6 ± 0.4 1.2 ± 0.2 0.870 ± 0.017 0.972 ± 0.007 ET Best baseline: 0.814 ± 0.020 7.1 ± 0.6 1.8 ± 0.3 0.806 ± 0.022 0.961 ± 0.009 Proposed: 0.835 ± 0.018 6.2 ± 0.5 1.5 ± 0.2 0.827 ± 0.020 0.965 ± 0.008 BraTS2020 WT Best baseline: 0.908 ± 0.012 4.9 ± 0.4 0.9 ± 0.1 0.898 ± 0.014 0.979 ± 0.006 Proposed: 0.918 ± 0.009 4.3 ± 0.3 0.7 ± 0.1 0.911 ± 0.011 0.982 ± 0.005 TC Best baseline: 0.870 ± 0.016 6.1 ± 0.5 1.3 ± 0.2 0.862 ± 0.018 0.970 ± 0.007 Proposed: 0.884 ± 0.014 5.4 ± 0.4 1.1 ± 0.1 0.877 ± 0.016 0.973 ± 0.007 ET Best baseline: 0.823 ± 0.019 6.7 ± 0.6 1.7 ± 0.2 0.815 ± 0.021 0.963 ± 0.009 Proposed: 0.842 ± 0.017 6.0 ± 0.5 1.4 ± 0.2 0.834 ± 0.019 0.967 ± 0.008 Open in a new tab Most notably, improvements are seen in the Enhancing Tumor that represents the most challenging subregion to segment. On BraTS2019, the Dice increased by over two points while the HD95 and ASSD rates have decreased significantly. On BraTS2020, we see the same patterns, which provides further evidence that the framework is highly effective in dealing with small, heterogeneous tumor regions. The improvements are also aligned with how the method was designed: (1) adaptive expert routing facilitates feature specialization in the presence of modality variability, (2) cross-modality consistency regularization provides additional rationality and alignment to representations, and (3) an uncertainty-aware decoder preserving the local boundary-aware training leads to sharper and more reliable contours. Statistical analysis of the qualitative difference in accuracy of Dice scores between the best performing baseline and other tested models show these results were statistically significant ( ) across all subregion of tumors. Despite the addition of new components such as adaptive expert routing and uncertainty-aware decoding to the proposed framework, there was a minimal increase in the number of parameters and FLOPs as compared to the previous version of the framework. By using LoRA, there was a substantial reduction in the number of trainable parameters relative to fully fine-tuned expert modules, resulting in a high degree of parameter efficiency. The additional amount of computational overhead was only approximately 1.3 million additional parameters and 9.2 GFLOPs, resulting in minimal increases in both inference time (0.04 seconds per volume) and memory usage (0.3 GB) as compared to the substantial improvements achieved for segmentation accuracy/boundary refinement. The computationally compared between proposed framework and best performing model is presented in Table 6 . Table 6. Computational complexity comparison between the proposed framework and the best-performing baseline model. Method Parameters (M) FLOPs (G) Inference time (s) Peak GPU memory (GB) Best baseline 36.9 182.4 0.78 ± 0.03 7.1 ± 0.2 Proposed (full) 38.2 191.6 0.82 ± 0.04 7.4 ± 0.2 Open in a new tab In addition, we analyzed inference time for the LoRA-MoE adaptive encoder (62%), whole decoder (27%), and uncertainty estimate (10% with 4 stochastic passes at ) to determine how suitable the system was for deployment. When using the deterministic single-pass mode, overall execution time has decreased from roughly 0.82 seconds/volume to approximately 0.74 seconds/volume while maintaining close to the baseline model execution time. Tables 7 and 8 shown below, demonstrate that the proposed methodology consistently improves the alignment of tumor sub-region boundaries across both the 2019 and 2020 BraTS datasets. In the 2019 dataset, the algorithm produces increases in the BF-score for WT (0.842 to 0.861), TC (0.801 to 0.823), and ET (0.756 to 0.779) indicative of increased contour precision and recall. The results for the 2020 BraTS dataset also reflect this improvement in BF-score across the nearly the same regions reinforcing the improvement of the structural consistency along the tumor boundary. Remarkably, the largest improvement was seen in the ET region, which is expected as it is naturally sensitive to boundary irregularities. In addition to this improvement, the previously reported reduction in HD95 and ASSD provide a stronger empirical basis for claiming that the incorporation of the boundary-aware supervision strategy leads to an improved and stable delineation of contours. Table 7. Boundary-focused performance comparison on the BraTS2019 dataset (mean ± std). Method WT BF-score TC BF-score ET BF-score Best baseline 0.842 ± 0.018 0.801 ± 0.021 0.756 ± 0.024 Proposed (full) 0.861 ± 0.015 0.823 ± 0.019 0.779 ± 0.022 Open in a new tab Table 8. Boundary-focused performance comparison on the BraTS2020 dataset (mean ± std). Method WT BF-score TC BF-score ET BF-score Best baseline 0.848 ± 0.017 0.808 ± 0.020 0.762 ± 0.023 Proposed (full) 0.869 ± 0.014 0.831 ± 0.018 0.786 ± 0.021 Open in a new tab The results in Table 9 indicate that our proposed framework greatly outperforms both baselines and ablative variants in all tumor subregions. The p-values for all comparisons were below the predefined significance threshold ( ), supporting the conclusion that our observed increases in Dice are not due purely to random variation across runs. Table 9. Statistical significance analysis (paired t -test p-values) comparing the proposed model with baselines and ablations on BraTS2019 (dice score). Comparison WT TC ET vs Best baseline 0.004 0.006 0.003 vs w/o CMC 0.002 0.005 0.001 vs w/o UAD 0.007 0.009 0.004 vs w/o BAT 0.006 0.008 0.005 Open in a new tab Robustness under varying missing-modality levels In order to re-evaluate robustness against modality absence at various levels, we analyzed the stratified performances based on which missing modalities were included in the analysis by way of three strata of missing modalities: (1) single modality removed; (2) two modalities removed; (3) three modalities removed. The performance of the proposed framework was measured across each strata group under the same training and test conditions, and the average Dice scores derived from all modality combinations within each of the three strata will be presented. This stratification provides a mechanism for evaluating increasingly degraded performance as a function of the missing modality. Table 10 displays that the segmentation performance decreased as the number of missing modalities increases. However, degradation remained relatively controlled and consistent, thereby indicating that the adaptive expert routing and cross-modality consistency mechanisms exhibit stability. Even when three modalities are removed in the analysis, the model continues to produce reasonable dice performance for all subregion of tumor, thereby demonstrating that the model has good resiliency to very incomplete input. The location that shows the greatest amount of performance decrease from the removal of three modalities is located within the ET Region due to its inherent reliance on contrast-specific information; however, the use of uncertainty-aware decoding and boundary guided supervision preserve the structural consistency within the proposed framework. Table 10. Stratified performance analysis under varying missing-modality levels on BraTS2019. Missing level WT dice TC dice ET dice Single-modality missing 0.905 ± 0.012 0.870 ± 0.016 0.828 ± 0.019 Double-modality missing 0.891 ± 0.014 0.856 ± 0.018 0.814 ± 0.021 Triple-modality missing 0.872 ± 0.017 0.839 ± 0.021 0.792 ± 0.024 Open in a new tab Tables 11 and 12 support This significant finding is rather surprising to see. Both data sets reported a removal of the T1ce modality has a greater impact on performance than the removal of any other modality especially in the ET region, this demonstrates the importance of using contrast enhanced images, evidence can also support that all combinations result in both a continuous decrease in accuracy and likewise for the standard deviation with respect to each configuration suggesting that performance is predictable even with many missing modalities. The findings in BraTS2019 and BraTS2020 reveal greater than or less than corresponding fluctuations for both data sets indicating equal robustness across both data sets. Testing was performed using only the 10 images from the BraTS2019 and BraTS2020 databases that had previously reported deterministic configuration-based measures rather than through the random discarding of information during the inference period. Table 11. Performance under full and missing-modality configurations on BraTS2019. Configuration WT TC ET Full modalities 0.918 ± 0.011 0.889 ± 0.014 0.852 ± 0.018 w/o T1 0.910 ± 0.012 0.882 ± 0.015 0.846 ± 0.019 w/o T1ce 0.892 ± 0.014 0.861 ± 0.018 0.803 ± 0.022 w/o T2 0.908 ± 0.013 0.878 ± 0.016 0.841 ± 0.020 w/o FLAIR 0.899 ± 0.015 0.870 ± 0.017 0.832 ± 0.021 Two modalities missing 0.891 ± 0.016 0.856 ± 0.019 0.814 ± 0.023 Three modalities missing 0.872 ± 0.018 0.839 ± 0.022 0.792 ± 0.026 Open in a new tab Table 12. Performance under full and missing-modality configurations on BraTS2020. Configuration WT TC ET Full modalities 0.924 ± 0.010 0.894 ± 0.013 0.861 ± 0.017 w/o T1 0.915 ± 0.011 0.887 ± 0.014 0.853 ± 0.018 w/o T1ce 0.898 ± 0.013 0.867 ± 0.017 0.812 ± 0.021 w/o T2 0.913 ± 0.012 0.883 ± 0.015 0.848 ± 0.019 w/o FLAIR 0.905 ± 0.014 0.875 ± 0.016 0.839 ± 0.020 Two modalities missing 0.897 ± 0.015 0.861 ± 0.018 0.821 ± 0.022 Three modalities missing 0.879 ± 0.017 0.844 ± 0.021 0.801 ± 0.025 Open in a new tab Qualitative results Besides numerical references, assessing a qualitative perspective of the segmentation masks can provide insights into the extent to which the proposed framework accurately separates tumor subregions across the varied cases from patients. Figures 3 and 4 provide relevant examples from the BraTS2019 and BraTS2020 datasets. Each row represents one subject slice and shows the FLAIR image, ground truth annotations, and the models’ predictions for the three clinically relevant tumor subregions: necrotic/core, edema, and enhancing tumor. Fig. 3. Open in a new tab Qualitative segmentation results on the BraTS2019 dataset. Fig. 4. Open in a new tab Qualitative segmentation results on the BraTS2020 dataset. The results for BraTS2019 as shown in Fig. 3 , indicate that the proposed model accurately identifies heterogeneous tumor structures, even in cases that are moderately challenging due to irregularity and diffuse boundaries. Example one depicts a lesion that has been defined, with apparent hyper-enhancement and edema, around the borders of enhancing regions. All three predicted classes spatially matched very well with the actual location of each class, with the tumor core accurately placed and enhancing regions having high agreement with respect to the respective edges. The only location where a slight blurring is evident is around the edges of the peripheral edema, but overall the contoured edges and morphology of the lesion is preserved. Example two depicts a lesion that has an irregular structure; however, the model accurately reconstructed the central core of the tumor and the enhancing regions of the tumor. Slight under-segmentation was present at the edges of the peripheral edema, likely due to the similarity in intensity between the edges of the edema and the surrounding tissue. Otherwise, overall tumor size and relative size of at least one internal sub-region was maintained with respect to ground truth data. In example three, the lesion occupied a fairly large area that involved most of the extent of one hemisphere of the brain. Both the predicted and the actual lesion contours for the enhancing region were similar for the most part. The transition of boundaries between the tumor core and the edema were also maintained with the majority of all contours and shapes having some degree of smoothing. In example four, there was limited lesion volume and although this may have caused the model to not detect or delineate the tumor, the model accurately located the lesion and has no false positives. The predicted core of the tumor and the predicted enhancing regions were also similar to ground truth data demonstrating robustness under very low tumor burden conditions. In example five, the tumor was a moderate size with unique enhancing characteristics. Although there were some discrepancies between the model repercussion of the enhancing region and the ground truth location of the enhancing region, the overall geometric shape of the lesion is similar regardless if using model versus ground truth data. The predictions on BraTS2020, as shown in Fig. 4 , further confirm the resiliency of the framework with missing modalities. Example one is a centrally located lesion that exhibits enhancement and edema. The predicted segmentation from the model for this example perfectly maps the enhancement and tumor core to the original, while the outer boundary of the entire tumor closely matches the original with minimal smoothing of the exterior of the whole tumor. Example two shows a heterogeneous tumor with varying levels of intensity and the multi-class predicted structures in the enhancing tumor regions correlate with the multi-class predicted structures in the tumor core. Although very small differences in the boundary were identified between the region of edema and the region of tumor, this occurred primarily in very low-contrast areas; as a result, the structural organization of the tumor was preserved. Example three is a small enhancement that extends only minimally into the surrounding area, therefore the enhancement and tumor core were accurately predicted. A minimal amount of under-segmentation was detected at the outer boundary of the tumor typical of an enhanced small volume example. Example four has an eccentric tumor that has irregular margins as it relates to the multi-class segmentation method and accurately reconstructs the shape of the tumor and its hierarchical sub-regions. Minor differences were found between the boundary of the tumor and the brain; however, the overall match between the original and prediction is substantial and perfect. Example five is an example of a challenging diffuse edema with slight enhancement. The proposed method maintains the spatial integrity of the whole tumor and the core of the tumor, while accurately predicting the location of the enhanced tumors. Minimal external smearing was found in each example; however, the overall topology of the tumor matched the original examples in cases presented. The framework that was put forward showed a more consistent and stable contour than was previously shown but did not account for some outliers. For example, with reduced shape types of a triple segmented area ET, there can be occasional under-segmentation due to a lack of contrast-type specific cues for the under-segmented; when irregular shapes of tumors are present or diffuse patterns, then some smoothing of the boundaries was also seen. This, in our opinion, could be due to the uncertainty-guided decoding was favoring structural stability over fidelity of local fluctuations. There were also rare false positive activations of of edema regions; this validates that improved consistency and adaptive routing promote stability, but still some level of segmentation performance is related to the discriminative type of modalities, specifically for enhancing finer detail. Ablation analysis To systematically evaluate the contribution of each module, we constructed the following ablation variants. Single-Expert: Replaces the mixture-of-experts routing mechanism with a single shared expert without dynamic gating. w/o CMC: Removes cross-modality consistency regularization while retaining adaptive routing and decoding. w/o UAD: Removes uncertainty-aware decoding and Monte Carlo sampling. w/o BAT: Removes boundary-aware supervision and curriculum scheduling. Full LoRA-MoE: Complete proposed framework with adaptive routing, LoRA insertion, CMC, UAD, and BAT. In addition, we evaluate a Full Fine-Tuning baseline where expert weights are fully updated without low-rank adaptation to assess computational trade-offs. The ablation results are shown in Table 13 . In conclusion the contribution of each component of the proposed pilot would be demonstrated in the overall metric we report. The full model achieves the best Dice score (0.875) and the lowest boundary errors (HD95 = 5.4, ASSD = 1.17) indicative of both proper volumetric overlap and appropriate boundaries. With the loss of cross-transmodal consistency (CMC) the Dice score drops and the boundary metrics also degrade which demonstrates the importance of the cross-model alignment effect. When the uncertainty-aware decoder (UAD) or boundary-aware training (BAT) were excluded there was IoU and boundary refinement we witnessed worse results as reflected in the HD95 and ASSD. Of note is that the curriculum learning (CL) provides stable optimization, whilst removing was a non-subtle drop in the Dice. The LoRA adaptation ensures computational efficiency without decreasing performance, recalling the reduced accuracy maintained better complexity of the overall model and accuracy when removed. The largest decline of a mean score was in the case of the experimental design of a single-Expert included where it was identified there was the lowest overall Dice (0.852) and worst boundary identified, indicating the clear benefit of a transmodal adaptive expert routing. In summary all contributions, indicate that each of the unique modules contributed meaningfully and their interactions in the end produced the robust segmentation of BraTS2019 we observed. Table 13. Extended ablation study on BraTS2019 dataset. Method Dice HD95 ASSD Proposed (full) 0.875 ± 0.012 5.4 ± 0.4 1.17 ± 0.1 w/o CMC 0.862 ± 0.013 6.1 ± 0.5 1.34 ± 0.1 w/o UAD 0.857 ± 0.015 6.5 ± 0.6 1.42 ± 0.2 w/o BAT 0.860 ± 0.014 6.6 ± 0.5 1.40 ± 0.2 w/o CL 0.864 ± 0.013 6.0 ± 0.5 1.32 ± 0.1 w/o LoRA 0.867 ± 0.012 5.9 ± 0.4 1.29 ± 0.1 Single-Expert 0.852 ± 0.016 6.8 ± 0.6 1.48 ± 0.2 Open in a new tab The ablation results in Table 14 further support the effectiveness of each of the modules proposed in this paper. The full model achieves the best Dice score (0.877) and lowest boundary errors (HD95 = 5.3, ASSD = 1.07), indicating the model correctly provides volumetric overlap as well as accurately defines the boundaries. Removing cross-modality consistency (CMC) shows a drop in Dice ( score along with substantially worse boundary metric degradation, which supports CMC’s role to stabilize feature representations when information is missing from a modality. Removing the uncertainty-aware decoder (UAD) or the boundary-aware training (BAT) will lead to the worst reductions in HD95 and ASSD. This shows that UAD or BAT are necessary for improvement of the tumor boundaries from segmentation masks. Removing curriculum learning (CL) produces scores for the Dice and other metrics to indicate reduce in performance, which supports CL’s role to stabilize optimizations for small or irregular-geometry tumor regions. LoRA adaptation once again exhibits the importance of UAD and CL as we see a drop-in performance with a slight increase in complexity. The single-expert variant exhibits the worst performance for all metrics, as Dice drops to 0.854 and yet boundary metrics are once again much worse, showing that expert routing is a key component to assist the model in handling the issues across heterogeneous modality scenarios. In general, when combined, these results support that including all modules is important to achieving reproducibly comparable performance on BraTS2020. Table 14. Extended ablation study on BraTS2020 dataset (mean ± std over five runs). Method Dice HD95 ASSD Proposed (full) 0.877 ± 0.011 5.3 ± 0.4 1.07 ± 0.1 w/o CMC 0.864 ± 0.013 6.0 ± 0.5 1.24 ± 0.1 w/o UAD 0.859 ± 0.014 6.4 ± 0.6 1.33 ± 0.2 w/o BAT 0.861 ± 0.014 6.5 ± 0.5 1.30 ± 0.2 w/o CL 0.865 ± 0.013 6.1 ± 0.5 1.26 ± 0.1 w/o LoRA 0.868 ± 0.012 5.8 ± 0.4 1.22 ± 0.1 Single-Expert 0.854 ± 0.015 6.7 ± 0.6 1.36 ± 0.2 Open in a new tab Discussion The findings on BraTS2019 and BraTS2020 support the overall utility of the suggested Adaptive LoRA-MoE framework for brain tumor segmentation tasks under missing modality conditions. The results from all trials indicated consistent superior Dice scores and lower boundary values in all tumor subregions when compared to state-of-the-art baseline methods, showing volumetric accuracy and contour accuracy. The largest improvements in Dice scores were seen in the Enhancing Tumor subregion, which is clinically relevant and is especially difficult to segment due to the small size and heterogeneous texture of the tumor. These outcomes suggest that the presented modules can support the model in extracting subtle patterns that are generally lost on traditional architectures. The ablation studies included in the methods section also supported the insight of the benefits of adding each model component. Eliminating the cross-modality consistency penalty produced a sharp decline in Dice, indicating its effect on aligning the features for each modality for stable overlaps. Removing either the uncertainty-aware decoder or boundary-aware training produced an obvious degradation in the HD95 and ASSD boundaries, indicating their contributions to boundary distinctions. Removing the curriculum learning component even made earlier, smaller tumor regions unstable relative to larger tumor regions, again indicating the positive contributions of incorporating additional structures. Lastly, while removing LoRA increased the complexity of the model, the performance was not improved for each segmentation measure. The metrics dropped considerably for the single-expert variant compared to all the presented metrics, indicating that adaptive expert routing is necessary for heterogeneous modality availability. All results suggest that each component offered a complimentary benefit and collectively, all components offered the most potent segmentation measure. Although improvements were made to the model’s performance in general, there were still cases of failure, especially when triple modal removals had occurred. In these cases, some relatively small regions of enhancement for the tumor had under-segmented the tumor likely because of having insufficient contrast-specific cues. In addition, while the uncertainty-aware decoder reduces over-confident predictions in ambiguous regions, the extreme modality sparseness could lead to the loss of discriminative feature richness. Consequently, the findings imply that even though our proposed framework enhances robustness, segmentation fidelity will remain partially reliant on the availability of contrast-enhanced information. The framework’s system elements provide different benefits; however, they also yield various shortcomings. For instance, the mixture-of-experts routing leads to a high degree of specialization but also adds a small amount of variability to routing between batches. Second, cross-modality consistency regularization (CMCR) will stabilize latent alignment; however, when used with too much regularization, some of the unique properties of each modality will be lost. Monte Carlo dropout allows you to gather an estimate for uncertainty; however, this method of sampling will take longer than if you used deterministic sampling during inference. Lastly, Curriculum Learning is meant to encourage stable optimization; however, by doing so, an extra scheduling hyperparameter must be created and could need to be adapted on other datasets. Collectively these trade-offs confirm that the framework gives emphasis to robustness and to the structure of the system rather than to the simplicity of the system’s design. Unlike imputation-based techniques, which generate incomplete MRI modalities before performing segmentation, the proposed Framework is designed to allow for modality-based representations of data to be learned directly through adaptive expert routing. This prevents the potential for errors generated due to the introduction of errors during the process of generative reconstruction. Additionally, unlike modality agnostic models, which share encoders and do not provide specialization between encoders and the corresponding segmentation, the Conditional Routing mechanism proposed in this Framework preserves the benefits of encoding using modality-specific transformations while enforcing latent alignment. Therefore, the proposed Framework performs at an intermediate methodological level between generative completion and uniform modality invariant modeling and attempts to balance specialization and generalization. From a clinical perspective, improved segmentation performance of the tumor core and enhancing region is most beneficial, as these are the subregions important for treatment planning and monitoring. Reliable segmentation delineation of the tumor boundaries not only aids in evaluating the radiological assessment, but also has the potential to reduce inter-observer variability among manual annotators. Further, various metrics of the proposed framework underwent evaluation across both BraTS datasets, suggesting it has potential in generalizing across other MRI cohorts and to similar clinical settings. For clinical environments, the design is still computationally feasible from a deployment point-of-view. A small increase in both parameters and FLOPs will have little impact on the delay of inference operations, while deterministic inference without uncertainty sampling has a runtime similar to the baseline architectures. Mixture-of-experts routing will only activate a subset of experts (top- k ), thus preventing full use of the computational resources. These factors provide evidence of potential integration into current or near-current radiology workflow for real-time or near real-time environments. Aside from the differences in modalities that are missing leading to variability in measurements from an MRI, variability in MRI measurements may also arise through the use of different scanner manufacturers and protocols, as well as across different magnetic field strengths (1.5T versus. 3T). Since the BraTS datasets include data from many different institutions and a range of acquisition methods, the current study did not implement a specific method of stratifying the results based on the scanner or the magnetic field strength; rather this could have been done through the use of an adaptive expert routing mechanism with feature specialization that provides a conditional feature specialization to allow for comparisons between modalities, and through the use of cross-modal consistency regularization to assist with aligning latent variables across modalities. More research is needed to determine how robust a model is to variations in the data acquired based on acquisition methodologies through controlled studies with respect to their consistency within their respective modality and across modalities. Despite showing improved performance, the proposed framework still has some limitations that need to be addressed. One limitation is that evaluation was conducted using carefully curated challenge datasets instead of evaluating using completely heterogeneous multi-institutional clinical cohorts which typically exhibit much larger variability in scanners. Another limitation of the proposed framework is that despite routing and consistency mechanisms improving stability, high contrast-dependent sub-regions are still impacted by extreme sparsification of modality during implementation. Lastly, while Monte Carlo sampling is used to measure uncertainty, it is also important to note that utilizing Monte Carlo sampling will add approximately a unique amount of computational burden to methodology; therefore, future investigations should explore adaptive routing regularization, domain adaptation methods and lightweight approximations to distributional uncertainty in order to further enhance the generalizability and efficiency for method deployments. Conclusion A novel Adaptive LoRA–MoE framework was proposed in this study with a unique multi-task, uncertainty-aware holistic decoding approach and boundary-guided training strategy for brain tumor segmentation with missing MRI modalities . By bringing together adaptive routing to experts, cross-modality consistency, uncertainty-aware decoding, and curriculum driven optimization, the proposed framework balance volumetric overlap accuracy while also respecting the importance of boundary accuracy. Extensive experiments conducted using the BraTS2019 and BraTS2020 datasets indicated promising improvements over existing baseline performance across each modality with especially considerable improvement seen in the challenging Enhancing Tumor subregion. Ablation studies further established that each component of the proposed framework has distinct but complementary gains and combinations of these components provide the best achievable reliability losses. The framework has a strong clinical rationale as stable and accurate boundary delineation of tumor subregions can assist radiologists in treatment planning by better supporting precise treatment methods, and more generally, reducing inter-observer variability. Additionally, the efficiency of the proposed method across modalities with unknown or available resources reflects practical feasibility for real-world clinical deployment, supported by controlled parameter growth, competitive inference time, and manageable memory consumption. Next steps include testing the model on multi-institutional datasets to determine reliability across various clinical practice distribution shifts in clinical settings, and pursuing domain adaptation techniques to assess changes in distributions due to multi-institutional practices, and finally, to improve the computational efficiency required to deploy these practices in timely environments. Furthermore, pairing the framework with explainability methods will allow the end-users to comprehend how the provided evidence-dependent automation is achieved and will maximize trust in the framework. Overall, the proposed methods presents a new opportunity towards an accurate automated clinical tool for brain tumor segmentation yielding promising results. Future work will also investigate integration with domain adaptation strategies and scanner-invariant representation learning to further enhance generalization across heterogeneous clinical environments. Acknowledgements This scientific paper is derived from a research grant funded by Taibah University, Madinah, Kingdom of Saudi Arabia—with grant number 1110-15-447. Author contributions A.G.M.A. and F.A. contributed equally to the conception and design of the study. A.G.M.A. developed the methodology, performed the experiments, and conducted the data analysis. F.A. supervised the study, provided critical insights, and reviewed the results. Both authors contributed to manuscript writing, revision, and approved the final version. Data availability The implementation of the work can be downloaded from: https://doi.org/10.5281/zenodo.18331077 Declarations Competing interests The authors declare no competing interests. Informed consent This study utilizes only publicly available, de-identified datasets. In accordance with institutional policies, the secondary analysis of such data does not constitute human subjects research and is therefore exempt from Institutional Review Board (IRB) review. Accordingly, no informed consent was required. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Nasir, I. M., Tehsin, S., Damaševičius, R. & Maskeliūnas, R. Integrating explanations into CNNs by adopting spiking attention block for skin cancer detection. Algorithms 17 , 557 (2024). [ Google Scholar ] 2. Mushtaq, I. et al. Customer prioritization for medical supply chain during COVID-19 pandemic. Comput. Mater. Contin. 70 , 59–72 (2021). [ Google Scholar ] 3. Oktay, O. et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018). 4. Zhou, Z., Siddiquee, M. M. R., Tajbakhsh, N. & Liang, J. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Trans. Med. Imaging 39 , 1856–1867 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Isensee, F., Jäger, P. F., Kohl, S. A., Petersen, J. & Maier-Hein, K. H. Automated design of deep learning methods for biomedical image segmentation. arXiv preprint arXiv:1904.08128 (2019). [ DOI ] [ PubMed ] 6. Tehsin, S. et al. Enhancing signature verification using triplet siamese similarity networks in digital documents. Mathematics 12 , 2757 (2024). [ Google Scholar ] 7. Yousafzai, S. N. et al. X-news dataset for online news categorization. Int. J. Intell. Comput. Cybern. 17 , 737–758 (2024). [ Google Scholar ] 8. Wang, H. et al. Mixed transformer u-net for medical image segmentation. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . 2390–2394 (IEEE, 2022). 9. Hatamizadeh, A. et al. Unetr: Transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 574–584 (2022). 10. Xie, Z., Miao, G. & Chang, H. MTSR: Mamba-transformer super-resolution model for hyperspectral images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. (2025). 11. Nasir, I. M., Alrasheedi, M. A. & Alreshidi, N. A. MFAN: Multi-feature attention network for breast cancer classification. Mathematics 12 , 3639 (2024). [ Google Scholar ] 12. Malik, D. S. et al. Block cipher nonlinear component generation via hybrid pseudo-random binary sequence for image encryption. Mathematics 12 , 2302 (2024). [ Google Scholar ] 13. Jabbar, A., Li, X. & Omar, B. A survey on generative adversarial networks: Variants, applications, and training. ACM Comput. Surv. (CSUR) 54 , 1–49 (2021). [ Google Scholar ] 14. Liu, R., Zuo, H., Lian, Z., Schuller, B. W. & Li, H. Contrastive learning based modality-invariant feature acquisition for robust multimodal emotion recognition with missing modalities. IEEE Trans. Affect. Comput. 15 , 1856–1873 (2024). [ Google Scholar ] 15. Havaei, M., Guizard, N., Chapados, N. & Bengio, Y. Hemis: Hetero-modal image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention . 469–477 (Springer, 2016). 16. Wu, R., Wang, H., Chen, H.-T. & Carneiro, G. Deep multimodal learning with missing modality: A survey. arXiv preprint arXiv:2409.07825 (2024). 17. Lai, B. et al. Hetero-modal learning and expansive consistency constraints for semi-supervised detection from multi-sequence data. In International Workshop on Machine Learning in Medical Imaging . 296–305 (Springer, 2021). 18. Wei, S., Luo, Y., Ma, X., Ren, P. & Luo, C. Msh-net: Modality-shared hallucination with joint adaptation distillation for remote sensing image classification using missing modalities. IEEE Trans. Geosci. Remote Sens. 61 , 1–15 (2023). [ Google Scholar ] 19. Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning . 1597–1607 (PmLR, 2020). 20. Chiou, C.-Y. et al. Hierarchical cross-consistency network based unsupervised domain adaptation for pathology whole slide image segmentation. IEEE Open J. Eng. Med. Biol. (2025). 21. Yousafzai, S. N., Nasir, I. M., Tehsin, S., Fitriyani, N. L. & Syafrudin, M. Fltrans-net: Transformer-based feature learning network for wheat head detection. Comput. Electron. Agric. 229 , 109706 (2025). [ Google Scholar ] 22. Wang, S. et al. Prototype knowledge distillation for medical segmentation with missing modality. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . 1–5 (IEEE, 2023). 23. Kendall, A. & Gal, Y. What uncertainties do we need in Bayesian deep learning for computer vision? Adv. Neural Inf. Process. Syst. 30 (2017). 24. Leibig, C., Allken, V., Ayhan, M. S., Berens, P. & Wahl, S. Leveraging uncertainty information from deep neural networks for disease detection. Sci. Rep. 7 , 1–14 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Dalca, A. V. et al. Unsupervised deep learning for Bayesian brain MRI segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention . 356–365 (Springer, 2019). [ DOI ] [ PMC free article ] [ PubMed ] 26. Karimi, D. & Salcudean, S. E. Reducing the Hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Trans. Med. Imaging 39 , 499–513 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 27. Yousafzai, S. N. et al. Multi-Stage Neural Network-Based Ensemble Learning Approach for Wheat Leaf Disease Classification. (IEEE Access, 2025). 28. Zhou, T. Boundary-aware and cross-modal fusion network for enhanced multi-modal brain tumor segmentation. Pattern Recognit. 165 , 111637 (2025). [ Google Scholar ] 29. Shaker, A. et al. Unetr++: Delving into efficient and accurate 3D medical image segmentation. IEEE Trans. Med. Imaging 43 , 3377–3390 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Bengio, Y., Louradour, J., Collobert, R. & Weston, J. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning . 41–48 (2009). 31. Liu, Z., Wei, J., Li, R. & Zhou, J. Learning multi-modal brain tumor segmentation from privileged semi-paired MRI images with curriculum disentanglement learning. Comput. Biol. Med. 159 , 106927 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 32. Kervadec, H., Dolz, J., Granger, E. & Ben Ayed, I. Curriculum semi-supervised segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention . 568–576 (Springer, 2019). 33. Abdelbaki, W. et al. Domain-adaptive MRI learning model for precision diagnosis of CNS tumors. Biomedicines 14 , 235 (2026). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Chartsias, A. et al. Factorised spatial representation learning: Application in semi-supervised myocardial segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention . 490–498 (Springer, 2018). 35. Ding, H., Zhang, X., Lu, W., Yuan, F. & Luo, H. Mmaformer: Multiscale modality-aware transformer for medical image segmentation. Electronics 13 , 4636 (2024). [ Google Scholar ] 36. Nasir, I. M., Alshaya, H., Tehsin, S. & Bouchelligua, W. Transformer-based foundation learning for robust and data-efficient skin disease imaging. Diagnostics 16 , 440 (2026). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Yousafzai, S. N. et al. A hybrid deep learning approach integrating CNN and transformer for lung cancer classification using CT scans. Sci. Rep. (2026). [ DOI ] [ PMC free article ] [ PubMed ] 38. Hu, E. J. et al. Lora: Low-rank adaptation of large language models. Iclr 1 , 3 (2022). [ Google Scholar ] 39. Shazeer, N. et al. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017). 40. Liu, J., He, J., Zhang, J., Ren, J. S. & Li, H. Efficientfcn: Holistically-guided decoding for semantic segmentation. In European Conference on Computer Vision . 1–17 (Springer, 2020). 41. Henry, T. et al. Brain tumor segmentation with self-ensembled, deeply-supervised 3D u-net neural networks: A brats 2020 challenge solution. In International MICCAI Brainlesion Workshop . 327–339 (Springer, 2020). 42. Liu, Z. et al. Deep learning based brain tumor segmentation: A survey. Complex Intell. Syst. 9 , 1001–1026 (2023). [ Google Scholar ] 43. Isensee, F. et al. nnu-net: Self-adapting framework for u-net-based medical image segmentation. arXiv preprint arXiv:1809.10486 (2018). 44. Qamar, S., Ahmad, P. & Shen, L. Hi-net: Hyperdense inception 3 D unet for brain tumor segmentation. In International MICCAI Brainlesion Workshop . 50–57 (Springer, 2020). 45. Jia, Q. & Shu, H. Bitr-unet: A CNN-transformer combined network for MRI brain tumor segmentation. In International MICCAI Brain Lesion Workshop . 3–14 (Springer, 2021). [ DOI ] [ PMC free article ] [ PubMed ] 46. Ahmad, P., Qamar, S., Shen, L. & Saeed, A. Context aware 3D unet for brain tumor segmentation. In International MICCAI Brainlesion Workshop . 207–218 (Springer, 2020). 47. Zhang, R. et al. Hmnet: Hierarchical multi-scale brain tumor segmentation network. J. Clin. Med. 12 , 538 (2023). [ DOI ] [ PMC free article ] [ 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 implementation of the work can be downloaded from: https://doi.org/10.5281/zenodo.18331077 Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.0 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases 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

• AI-Based Methods in Neuropathology for Diagnosis and Treatment of Brain Tumors.#121121
• MobileNetV1-Based Deep Learning Model for Accurate Brain Tumor Classification#223198
• Semi-supervised Medical Image Segmentation via Perturbation-Aware Mutual Learning and Edge-Aware Uncertainty Loss for Accurate Anatomical Delineation#445028
• A Maze-Based Computational Framework for Tumor Localization in the Brain#698942
• Diffusion-tensor Imaging in Brain Tumors Evaluation#902157