Interactive AI assisted pediatric burn assessment based on smartphone images.

Interactive AI assisted pediatric burn assessment based on smartphone images - 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 12;16:16996. doi: 10.1038/s41598-026-48169-z Interactive AI assisted pediatric burn assessment based on smartphone images Hao Wang 1 , Shuaidan Zeng 2 , Weiqing Li 2 , Xiaodi Chen 2 , Qianqian Mei 2 , Huating Chen 2 , Lina Fu 2 , Zhenhui Zhao 2 , Shengping Tang 2 , Kaize Zheng 2, ✉ , Yanyan Liang 1, ✉ , Zhu Xiong 1, 2, ✉ Author information Article notes Copyright and License information 1 School of Computer Science and Engineering, Faculty of Innovation Engineering, Macau University of Science and Technology, Macao, 999078 China 2 Department of Pediatric Orthopedics, Shenzhen Pediatrics Institute of Shantou University Medical College, Shenzhen, 518034 China ✉ Corresponding author. Received 2025 Nov 29; Accepted 2026 Apr 6; 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: PMC13230800  PMID: 41968181 Abstract Burn injuries are a common pediatric health threat with depth assessment relying heavily on subjective visual inspection. While objective techniques like laser Doppler imaging exist, their cost and portability limitations restrict use. We propose SAM-DR to address the challenge of scarce annotated burn data by repurposing pre-trained models with minimal fine-tuning. By replacing SAM’s segmentation head with dense linear regression, our method not only identifies burn locations but also perceives burn depth through continuous depth prediction. Using 294 smartphone images from 94 patients annotated by 9 clinicians, we conducted a pixel-level comparison of human disagreement. SAM-DR achieved a 0.96 Dice score in wound segmentation, establishing state-of-the-art performance, and the use of interactive thresholding enabled segmentation of different burn depths comparable to human experts, suitable for assisted annotation. We developed an interactive tool based on SAM-DR that supports both clinical diagnosis and data annotation, offering a non-contact solution for burn assessment and dataset creation. Keywords: Burn segmentation, Burn depth identification, Interactive AI, Pediatric burns Subject terms: Computational biology and bioinformatics, Diseases, Engineering, Health care, Mathematics and computing, Medical research Introduction Burn injuries, particularly pediatric burns, represent a major global public health problem, often leading to long-term physiological dysfunction and psychological trauma 1 . In clinical practice, the depth of the burn is one of the most critical factors in determining treatment plans (e.g. conservative management or surgical intervention) and predicting patient outcomes 2 . However, current clinical practice relies heavily on the visual inspection and subjective experience of burn specialists. This diagnostic approach has been shown to produce significant variability between raters due to differences in physicians’ level of expertise and experience 3 . Such discrepancies are particularly pronounced in borderline cases, such as between deep partial-thickness and full-thickness burns, where the visual boundaries between different burn depths are often ambiguous and atypical. To pursue more objective and accurate assessment, advanced techniques such as Laser Doppler Imaging (LDI) have been developed, aiding depth assessment by evaluating micro-circulatory perfusion status in burn wounds 4 . Although its effectiveness has been validated, the high cost, operational complexity, and lack of portability of LDI equipment hinder its widespread adoption in resource-limited healthcare settings and many developing regions 5 . Consequently, developing an objective diagnostic tool that does not rely on specialized equipment and can be widely used has become an urgent need. In recent years, the combination of smartphone-based photography and artificial intelligence (AI) technology has offered a highly promising solution to this challenge 6 . The global ubiquity of smartphones makes them a low-cost and readily available imaging tool. Simultaneously, deep learning, particularly in the field of image segmentation, has made remarkable progress 7 . Previous studies have attempted to use Convolutional Neural Networks (CNNs) for burn area segmentation or overall burn depth classification 8 – 19 . However, these methods typically operate on pre-cropped wound areas or provide a general classification for a region, and have not yet achieved precise pixel-level segmentation that simultaneously discriminates between multiple burn depths. This finer-grained capability is crucial for enabling automated calculation of burn area and for mapping the heterogeneous depth distribution within a complex wound, which are key to comprehensive assessment. Recently, the Segment Anything Model (SAM) emerged as a foundational vision model, demonstrating unprecedented zero-shot generalization capability and flexible interactivity for natural image segmentation 20 . An effort named MedSAM attempted to adapt SAM’s capabilities to the medical imaging domain and achieved excellent performance on many modalities by fine-tuning on a large-scale medical image dataset 21 . However, while burn images fundamentally belong to the domain of natural images, they represent a specialized and relatively uncommon category within it. Their unique texture, color, and morphological characteristics differ significantly from both common natural scenes and typical medical images (e.g., X-rays, CT). Our preliminary tests indicated that MedSAM’s segmentation accuracy on burn images was even inferior to the original SAM, suggesting that fine-tuning for a broad spectrum of medical imaging might impair the model’s performance on such specialized natural images. While one previous attempt combined vanilla SAM with a detection model (e.g., Grounding DINO) for burn segmentation, that study lacked rigorous quantitative evaluation, leaving its clinical utility in question 11 . In terms of depth-specific segmentation, existing approaches have primarily relied on semantic segmentation frameworks. These typically formulate the task either as a dense multi-class classification problem to output a single mask with categorical labels 22 – 24 , or as multiple, independent binary segmentation tasks for each burn depth 25 . Notably, Rozo et al. employed an ordinal ranking approach for three-level depth assessment 26 . However, these frameworks are constrained by their reliance on discrete classification boundaries, which may not faithfully represent the continuous and often ambiguous transitions between burn degrees. In contrast, our SAM-DR (SAM-Depth Regression) method introduces a novel paradigm by performing dense depth regression at the pixel level. This approach fundamentally models the burn depth as a continuous value. The subsequent application of tunable thresholds to the regression output enables a flexible and highly accurate implementation of multi-depth segmentation, effectively bridging the gap between continuous depth estimation and discrete clinical categorization. Addressing these gaps, this study introduces SAM-DR, a novel interactive AI-assisted method for multi-depth segmentation and assessment. Our core contributions are threefold: We collected and annotated a dataset of pediatric burn images captured by smartphones in a real clinical environment. We engaged nine annotators across different seniority levels (including expert-level surgeons, attending physicians, and resident physicians) for independent annotation, enabling a pioneering quantitative analysis of pixel-level inter-physician annotation disagreement. We introduced a modification to the SAM architecture by replacing its mask binary classification head with a dense linear regression head. This enables the model to predict a continuous depth value for each pixel, representing burn severity on a continuous scale. The resulting depth map is then converted into discrete multi-depth segmentation masks through a flexible thresholding process, effectively bridging continuous depth estimation with clinical classification needs. We developed an interactive web-based tool. This tool not only provides clinicians with visualized depth heatmaps for diagnostic assistance but also generates high-quality pre-annotations by incorporating an interactive manual threshold, significantly improving the efficiency of future large-scale data annotation. Our codes and finetuned model will be share on github.com/HLSS-Hen/Interactive-AI-assisted-pediatric-burn-assessment-based-on-smartphone-images . Methods The overall pipeline of our proposed SAM-DR framework is illustrated in Fig. 1 , encompassing data curation, model adaptation, and evaluation. The following sections detail each component, including data collection and annotation, architectural modifications to the base SAM model, implementation specifics for training and inference, and the development of an interactive tool for clinical and research use. Fig. 1. Open in a new tab Overview of training and evaluation pipelines.. Ethics approval and consent to participate This study has been approved by the Medical Ethics Committee of the Shenzhen Children’s Hospital (approval No.:202505902). Written informed consent was obtained from all participants or their legal guardians before data collection and analysis. All procedures were in accordance with institutional and national ethical standards and the Declaration of Helsinki. Data collection A total of 294 smartphone images were collected from 94 hospitalized pediatric patients at Shenzhen Children’s Hospital, captured using various smartphone models. All images were captured during the debridement procedure approximately 48 hours post-burn, which represents the clinically accepted optimal timing for burn depth assessment 27 . To ensure data independence and avoid redundancy, images from the same patient that depicted overlapping or highly similar burn regions were carefully reviewed, and only one representative image was retained in such cases. To protect patient privacy, the rulers visible in the images, which contained patient information, were replaced with uniform color patches using Photoshop software prior to analysis. The dataset intentionally included challenging cases such as burns located on complex regions, atypical burn presentations, and images with suboptimal focus or lighting conditions to enhance model generalizability. Annotation process and ground truth generation In this study, burn regions were categorized into three distinct depth levels based on clinical consensus: Region 1#: Superficial partial-thickness burns and deeper, representing the entire wound area. Region 2#: Deep partial-thickness burns and deeper, indicating areas with scarring risk. Region 3#: Full-thickness burns. Nine clinicians with different expertise levels participated in the annotation process, including three expert-level surgeons, three attending physicians, and three resident physicians. Using the X-AnyLabeling tool 28 , each annotator independently delineated burn regions using polygons. We established two versions of ground truth: GT v1 represented the consensus between two in-service experts, who were selected as they are the primary clinicians responsible for actual diagnostic decisions within the clinical team, while the third expert had been retired for a significant period. The consensus was obtained by averaging and binarizing their annotations. GT v2 was created by an author (H.W.) with extensive annotation experience, who refined GT v1 annotations to improve boundary precision and cross-view consistency, with all modifications reviewed and approved by the original experts. The dataset was split at the patient level in an approximately 8:2 ratio, maintaining balanced representation of different burn depths across splits. The detailed distribution of patients and images across training and test sets is provided in Table 1 . Table 1. Distribution of images across training and test sets. Region Train set Test set Total 1# 60 15 75 2# 157 37 194 3# 20 5 25 Total 237 57 294 Open in a new tab Analysis of human disagreement We conducted a comprehensive analysis of inter-rater variability by calculating pairwise Dice coefficients among all 11 annotation sets (9 clinicians plus GT v1 and GT v2). To thoroughly investigate the nature of human disagreement, we performed two distinct types of comparisons: Hard Comparison: This evaluation followed the standard segmentation protocol, comparing the entire annotated areas for each burn depth using Dice coefficients. This approach primarily reflects disagreements in the overall identification and extent of burn regions. Soft Comparison: This more nuanced analysis was designed to isolate disagreements specifically along the boundaries of mutually acknowledged burn areas. For a given depth category and a pair of annotations A and B, we first identified their overlapping region C. Then, for each annotation, we performed instance segmentation based on pixel connectivity. Only the instances that intersected with the overlapping region C were retained and merged to form new binary masks for comparison. This soft comparison effectively focuses on the precision of boundary delineation, excluding disagreements arising from one rater completely missing a region annotated by another. This two-fold pixel-level disagreement analysis established a crucial human performance baseline for evaluating our AI model. SAM-depth regression (SAM-DR) method architecture Segment anything model (SAM) SAM comprises three main components: 1) An image encoder based on Vision Transformer (ViT) that extracts image embeddings; 2) A prompt encoder that processes user inputs (points, boxes and mask); and 3) A mask decoder that generates segmentation masks by combining image and prompt embeddings. The complete workflow includes preprocessing steps where input images are resized and normalized to a fixed size, and postprocessing steps where predicted masks are rescaled back to the original image dimensions. The detailed architecture of the original SAM model is illustrated in Fig. 2 . Fig. 2. Open in a new tab Architecture of the segment anything model (SAM). The encoding of sparse prompts is detailed as follows: Point embeddings are generated by adding a sinusoidal positional encoding of the point coordinates to a learned embedding weight (for foreground/background points), with padded points set to a learned no-point embedding. Box embeddings are formed by encoding the two corner coordinates via the same positional encoding, then adding the top-left corner’s encoding to one learned embedding weight and the bottom-right corner’s encoding to another. Preprocessing involves resizing the input image while maintaining its aspect ratio, with the longer side scaled to 1024 pixels. The resized image is then normalized and padded to form a square before being fed into the image encoder. The image encoder employs a standard Vision Transformer (ViT) architecture. It first splits the input image into patches and encodes them with positional information. These patches are then processed through successive transformer blocks to produce a dense feature map that captures comprehensive image context, forming the visual foundation for segmentation. Finally, the features are projected through a convolutional neck network to an optimal dimension for mask decoding. The prompt encoder encodes user interactions into prompt embeddings. Spatial prompts (points, boxes) are encoded using sinusoidal positional encoding combined with learned embeddings that distinguish prompt type (e.g., foreground/background point). Mask inputs are processed through a small convolutional network to match the image embedding dimensions. The mask decoder employs a lightweight transformer to fuse image features with prompt embeddings. It initializes learned tokens for mask prediction and IoU scoring, which interact with the visual features through cross-attention. The decoded mask tokens are processed by a transposed convolutional network to upsample the spatial resolution of the feature maps. Each mask token is then passed through an Multilayer Perceptron (MLP) to generate dynamic convolution kernels, which are applied to the upsampled features to produce the final masks. The IoU token is independently fed into another MLP to estimate mask quality. The architecture contains multiple task heads (1+3) operating in two modes: single-mask prediction or multi-mask output. In multi-mask mode, three separate mask-IoU pairs are generated, functioning as an expert ensemble where the highest-IoU prediction can be selected. Postprocessing converts the generated masks back to the original image size through a two-step procedure: first upsampling the mask to using bilinear interpolation, then cropping to the actual input dimensions and resizing to match the original image resolution. Architectural modification To preserve the powerful pre-trained capabilities of the original SAM model while adapting it for burn depth estimation, we introduced a targeted modification by replacing the mask classification head with a dense linear regression head (a change that does not require structural alterations to the model). This adaptation transforms the task from pixel-wise binary classification to continuous depth regression, enabling direct prediction of burn depth values. Between the mask decoder and postprocessing stages, we apply a sigmoid activation to the model output and multiply by 3, mapping the regression values to a 0-3 range. This scaling corresponds to clinical categories: 0 for non-burn regions, and 1, 2, 3 for burn regions 1#, 2#, and 3#, respectively. Accordingly, the loss function was switched from the combination of Focal Loss and Dice Loss to Mean Squared Error Loss. We froze both the point/mask prompt processing modules and the multi-mask prediction heads, as empirical tests confirmed that enabling the multi-mask prediction heads provided no performance improvement in our specific scenario. Integration of interactive thresholding The dense linear regression output of our SAM-DR model produces continuous depth values, requiring thresholding to obtain discrete segmentation masks for quantitative evaluation. While empirical global thresholds provide a standardized conversion method suitable for automated clinical assessment and represent the lower bound of the model’s capability, we further enhanced the framework’s flexibility by introducing an interactive thresholding mechanism, which demonstrates the upper bound of performance achievable through optimized threshold selection. This interactive approach is particularly valuable for data annotation purposes, where it significantly improves efficiency. During the pre-annotation process, annotators can manually adjust a global threshold through an intuitive interface to achieve optimal segmentation quality. The process requires only two clicks to define the region of interest followed by threshold adjustment, enabling rapid generation of high-quality pre-annotations. The performance achieved through this manual adjustment demonstrates the upper bound of our regression model’s capability with optimized thresholding. This dual-threshold strategy provides both standardized automated processing for clinical applications and enhanced flexibility for research and annotation purposes, making our framework adaptable to diverse usage scenarios. Implementation details Figure 1 illustrates the comprehensive pipeline of our proposed method, encompassing the entire workflow from data preparation and model training to evaluation. The following sections provide detailed descriptions of each component in this pipeline. Data augmentation We employed a sequential data augmentation strategy including random cropping (preserving all annotated regions), random flipping, and RandAugment 29 . The RandAugment transformations consisted of Identity, Rotate, Brightness, Color, Contrast, Sharpness, Posterize, AutoContrast, and Equalize operations, while excluding geometric transformations such as shear and translation. During the AutoAugment process, two augmentation operations were randomly selected from this set and applied sequentially to each image. Prompt sampling strategy To effectively simulate real-world clinical usage patterns during training, we developed a comprehensive prompt sampling strategy. For each training iteration, we generated three types of prompts corresponding to distinct clinical operational modes: Boxes Mode: Local bounding boxes computed as the minimum enclosing rectangles for each isolated burn region (based on pixel connectivity), obtained through connected component analysis on the annotated binary masks. This simulates detailed region-specific annotation. One Box Mode: A global bounding box encompassing all burn regions within an image, simulating scenarios where clinicians provide a coarse region of interest for comprehensive assessment. None Box Mode: A ”none” prompt condition that enables fully autonomous inference without user guidance, representing rapid screening applications. This multi-faceted approach ensures robust model performance across various clinical interaction paradigms, from detailed annotation to rapid screening scenarios. Benchmark methods and training configuration To ensure comprehensive performance evaluation, we implemented and compared several established segmentation architectures, including U-Net 7 and DeepLabV3Plus-ResNet101 30 . The DeepLabV3Plus-ResNet101 model was initialized with weights pre-trained on the Pascal VOC 2012 dataset 31 . For each model architecture, we systematically evaluated both semantic segmentation and dense depth regression approaches, adapting the final output layers accordingly to accommodate each task’s requirements. For the SAM-based semantic segmentation (SAM-SS), we combined the 1+3 output masks to form a four-class semantic segmentation output (background, superficial partial-thickness, deep partial-thickness, and full-thickness burns). To address class imbalance in the training data, we implemented an online random oversampling strategy. During each training epoch, the training set was independently resampled to 480 images (160 images per class), with oversampled images randomly selected from the original data and augmented online. This per-epoch sampling ensured that the model encountered varied data distributions throughout training, enhancing generalization while maintaining class balance. For semantic segmentation models, we employed a combined loss function of cross-entropy loss and Dice Loss. Specifically, we used the standard cross-entropy loss, which corresponds to the Focal Loss with . The Dice Loss is computed as , where and represent the predicted and ground truth values respectively, and is a smoothing factor. For depth regression models, we used MSE Loss: , where is the ground truth depth value and is the predicted value. All benchmark models underwent extensive training for 120 epochs to ensure complete convergence. In contrast, the SAM-based methods, by fully leveraging the powerful pre-training, achieved optimal performance with only 30 epochs of fine-tuning. This significant reduction in required training epochs demonstrates the substantial efficiency advantages of building upon foundational models with extensive pre-training. All experiments were conducted on an NVIDIA RTX 3090 GPU using PyTorch 2.8, Transformers 4.51.3, OpenCV-Python 4.12.0.88 and CUDA 12.6. Evaluation metrics We adopted the Dice coefficient as the primary quantitative metric for evaluating segmentation performance, given its prevalence in medical image analysis and direct relationship with Intersection over Union (IoU), where . For all semantic segmentation methods, the evaluation was conducted by re-categorizing the predicted clinical burn categories into regions 1#, 2#, and 3# for comparison with the corresponding ground truth regions, rather than directly comparing the clinical categories. This approach ensured a consistent evaluation framework across different method paradigms. For depth regression models, which require thresholding to convert continuous values to discrete segments, we evaluated two approaches: global fixed threshold and optimal threshold. The optimal threshold approach was used to simulate the upper-bound performance achievable through manual threshold selection in an automated manner. Comprehensive evaluation was conducted across three operational modes: Boxes Mode (individual prompts for each burn region), One Box Mode (single prompt covering all affected areas), and None Box Mode (fully autonomous operation), providing insights into model performance across different levels of clinical guidance, from detailed annotation to rapid screening scenarios. Tool development We developed an interactive web application using Gradio to support clinical assessment and data annotation. The tool automatically loads burn images from a specified directory and also allows users to upload images in real-time during clinical sessions, with all images stored for subsequent analysis. The system provides immediate visualization of depth heatmaps and distribution histograms using our model’s prompt-free inference mode. The interface enables two main interaction types: box prompts created through simple two-click operations for localized analysis, and interactive threshold adjustment via histogram manipulation for converting continuous depth values to discrete segmentations. The system incorporates contour detection and simplification algorithms (OpenCV’s findContours function with Douglas-Peucker approximation) to generate polygon annotations with optimized vertex counts. All generated annotations can be exported in LabelMe-compatible format, ensuring seamless integration with existing annotation workflows. This integrated approach provides a unified solution for both clinical reference and efficient data pre-annotation. Results Dataset characteristics A total of 294 smartphone images were collected from 94 pediatric patients (65 males and 29 females, ratio 1.85 : 1) with a mean age of years (mean±standard deviation). The images were acquired using various smartphone models, yielding a diverse set of resolutions. The image dimensions ranged from (1.55 megapixels to (12.19 megapixels), with a median dimension of (3.76 megapixels). The mean total pixel count was megapixels. All subsequent analyses–including human disagreement assessment and model evaluation–were performed at the original image resolution to maintain clinical realism. The anatomical distribution of burn regions is summarized in Table 2 and illustrated in Fig. 3 , which correlates patient age, burn region, and gender. Table 2. Distribution of burn regions in the training and test sets. Region Train set Test set Total Head 17 6 23 Trunk 39 6 45 Upper Limb 37 12 49 Lower Limb 33 11 44 Open in a new tab Fig. 3. Open in a new tab Distribution of burn characteristics across the pediatric cohort. Central scatter plot shows age versus burn regions, colored by gender. Top and right panels display age and regional frequency distributions, respectively. Analysis of human disagreement in visual assessment A comprehensive panel analysis of inter-rater variability in burn depth annotation is presented in Fig. 4 . Pairwise Dice coefficients were calculated across all annotators and three burn depth regions using two complementary methods. The hard comparison (top row), based on standard semantic segmentation evaluation, revealed disagreements in region identification and depth classification. The soft comparison (bottom row), designed to isolate boundary delineation differences in mutually acknowledged areas, showed substantially higher agreement levels. These quantitative results demonstrate significant variability in clinical burn assessment, particularly in boundary delineation and depth classification. Fig. 4. Open in a new tab Inter-rater variability in burn depth annotation. Pairwise Dice coefficients under hard (top row) and soft (bottom row) comparisons for region 1# (left), 2# (middle), and 3# (right). Annotators 0-2: expert-level; 3-5: attending; 6-8: resident; 9: GT v1; 10: GT v2. Hard comparison evaluates overall region identification, while soft comparison focuses on boundary delineation. To establish performance baselines for AI model evaluation, we computed the average Dice scores of all human annotators against GT v2 under both comparison methods. Under hard comparison, the mean Dice scores were for region 1#, for region 2#, and for region 3#. The corresponding soft comparison results showed improved agreement: for region 1#, for region 2#, and for region 3#. These results demonstrate significant variability in clinical burn assessment, with hard comparison reflecting both region identification and depth classification disagreements, while soft comparison primarily captures boundary delineation differences. The established human performance benchmarks provide meaningful reference points for subsequent model evaluation. Performance comparison of segmentation methods Table 3 presents a comprehensive comparison of segmentation performance between semantic segmentation (SS) and depth regression (DR) approaches across different model architectures, where all DR models utilized a standardized fixed thresholding scheme (0.5/1.5/2.5) for discrete segmentation with minimal hyperparameter tuning. This consistent threshold selection ensures fair comparison across all models and eliminates selection bias. Quantitative results demonstrate the consistent superiority of the DR paradigm over SS across all models for regions 1# and 2#, with several models approaching or exceeding human performance levels. Table 3. Performance comparison of different methods on burn wound segmentation. SS means semantic segmentation and DR means depth regression. Results are presented as Dice coefficients (mean ± standard deviation for human performance; Boxes Mode / One Box Mode / None Box Mode for AI models). Human performance is evaluated under two comparison paradigms: Hard comparison reflects disagreements in both region identification and depth classification, while Soft comparison isolates boundary delineation differences in mutually acknowledged areas. SAM-DR demonstrates superior performance across all regions and operational modes. Method Model Dice 1# 2# Human Hard Soft Pre-train SAM(ViT-b) / / - - MedSAM(ViT-b) / / - - SS U-Net - / - / - / - / DeepLabV3P(ResNet-101) - / - / - / - / SAM-SS(ViT-b) / / / / DR U-Net - / - / - / - / DeepLabV3P(ResNet-101) - / - / - / - / SAM-DR(ViT-b) / / / / Open in a new tab For U-Net architecture, the DR approach achieved Dice scores of and for regions 1# and 2# respectively, representing improvements of 0.038 and 0.093 over the SS variant ( , ). DeepLabV3Plus with DR obtained scores of and , showing consistent improvements over its SS counterpart ( , ). Our proposed SAM-DR method with fixed thresholding achieved the highest performance across all operational modes. In Boxes Mode, SAM-DR attained Dice scores of and for regions 1# and 2#, demonstrating statistically significant improvement over SAM-SS ( , ) in region 1# (paired t-test: t(56) = 3.049, p = 0.0035, Cohen’s d = 0.298). The performance advantage was maintained in One Box Mode, where SAM-DR achieved / compared to SAM-SS’s / , with region 1# showing significant improvement (t(56) = 2.497, p = 0.0155, Cohen’s d = 0.229). In None Box Mode, SAM-DR achieved / compared to SAM-SS’s / , though the difference in region 1# did not reach statistical significance (t(56) = 1.174, p = 0.2455, Cohen’s d = 0.096). This consistent performance pattern across all operational modes highlights the robustness of the depth regression approach. Due to the limited number of full-thickness burn cases in our dataset, all models showed limited capability in identifying region 3# burns. Although SAM-SS achieved a higher Dice score (0.20-0.23) for region 3# compared to other methods, this result is considered unreliable given the small sample size and substantial variability in predictions. Figure 5 provides qualitative comparisons of SS and DR outputs on representative test samples. The DR approach demonstrates superior boundary delineation and better handling of ambiguous transition zones between burn depths. In complex cases involving non-typical burn appearances and challenging regions, DR maintains more consistent performance compared to SS. Fig. 5. Open in a new tab Qualitative comparison of segmentation results. The figure shows representative outputs from three models (U-Net, DeepLabV3Plus, SAM) under two approaches: semantic segmentation (SS) and depth regression (DR). For DR, both the raw regression output and the manually thresholded result (MT) are displayed. Effect of interactive thresholding Quantitative evaluation of interactive threshold adjustment demonstrates substantial performance improvements over fixed-threshold approaches (Table 4 ). Manual thresholding enhanced segmentation accuracy across all models, with several exceeding human performance in the soft comparison paradigm. Table 4. Performance comparison of depth regression methods with fixed versus manual thresholding. Results are presented as Dice coefficients (mean ± standard deviation for human performance; Boxes Mode / One Box Mode / None Box Mode for AI models). Human Soft comparison represents boundary delineation performance in mutually acknowledged areas. Method Model Dice 2# 3# Human Soft Depth regression (fixed threshold) U-Net - / - / - / - / DeepLabV3P(ResNet-101) - / - / - / - / SAM-DR(ViT-b) / / / / Depth regression (manual threshold) U-Net - / - / - / - / DeepLabV3P(ResNet-101) - / - / - / - / SAM-DR(ViT-b) / / / / Open in a new tab U-Net with manual thresholding achieves Dice scores of , , and for regions 1#, 2#, and 3# respectively, representing improvements of 0.029, 0.254, and 0.369 over fixed-threshold DR outputs. DeepLabV3Plus attains , , and , showing improvements of 0.235 and 0.381 respectively in region 2# and region 3#. Our SAM-DR model with manual thresholding achieves the highest performance metrics, with Dice scores of , , and for regions 1#, 2#, and 3# in Boxes Mode. These results surpass the fixed-threshold DR outputs by 0.265 and 0.576 respectively in region 2# and region 3#, and exceed human soft comparison performance across all three regions. The performance advantage is consistently maintained across all operational modes, with Single Box Mode ( , , ) and None Box Mode ( , , ) demonstrating robust segmentation capability. Application in clinical practice and data annotation The proposed interactive tool was evaluated by all annotators participating in this study. Unanimous feedback indicated that the tool made high-precision annotation both easier and faster to accomplish. The tool’s interface and operational workflow are presented in Fig. 6 . The system provides immediate visual feedback through depth heatmaps and enables real-time prediction updates, facilitating efficient assessment of burn regions and serving as a valuable reference in clinical evaluation. Fig. 6. Open in a new tab Tool interface for burn assessment and annotation. ( a ) Automatic visualization of the depth heatmap and distribution histogram upon image loading. ( b ) Localized analysis of the entire wound area via a bounding box prompt, yielding a display similar to the previous subfigure due to comprehensive region selection. ( c ) Analysis of a sub-region, where the depth histogram shows distinct characteristics, and interactive thresholding converts the heatmap to segmented polygons. ( d ) Multi-region annotation result, where users can click the ’Save Annotations’ button to export the annotation file. The tool demonstrates notable flexibility across operational modes, supporting both detailed region-specific analysis through Boxes Mode and rapid screening via None Box Mode, making it adaptable to diverse clinical scenarios. Furthermore, the seamless integration with established annotation workflows through LabelMe-compatible export functionality enhances its practical utility in clinical and research settings. Discussion This study systematically investigates an interactive AI-based approach for burn multi-depth segmentation, achieving significant progress across multiple dimensions. Compared to previous research, our work introduces dense depth regression to the field of burn segmentation, breaking through the limitations of traditional classification or detection paradigms. In recent years, automated burn assessment research can be categorized into three types based on the task: classification-based methods predicting overall burn depth from wound images 8 – 18 , detection-based frameworks identifying burn regions 14 , and segmentation-based methods focusing on burn area identification, with the best reported performance around 0.92 Dice 15 – 18 , 25 . Notably, some studies have begun exploring more specialized application scenarios, such as predicting areas failing to heal within 21 days 19 or calculating TBSA via palm segmentation 25 . Rozo et al. utilized LDI device diagnostic results as targets, implementing three-depth analysis using ordinal classification 26 ; while Zhang et al. employed semantic segmentation for three-depth burn analysis, noting that the scarcity and difficulty of annotating burn data limit the performance of traditional deep learning-based methods 24 . Regarding multi-depth segmentation, the work by Chang et al. is closest to our study, achieving two-level segmentation of the burn area and potentially poorly perfused deep burn areas with Dice scores of 0.906 and 0.897, respectively 25 . In contrast, our SAM-DR method achieves state-of-the-art burn wound segmentation while simultaneously providing more burn depth discrimination capability. This study quantifies the disagreement among clinicians in burn depth assessment at pixel-level precision, offering a new perspective for understanding this clinical challenge. Hard comparison results show that relatively high agreement (Dice ) was achieved among clinicians for region 1# (wound area), with disagreements mainly reflected in the annotation precision of wound edges. However, for region 2# (areas with scarring risk), the Dice coefficient among clinicians dropped significantly to . This notable decline primarily stems from two factors: the inherently ambiguous boundaries between different burn depths, and substantial disagreements in depth judgment among clinicians–some annotated region 2# within region 1# areas, while others did not annotate corresponding areas. The situation is more complex for region 3#, with a Dice coefficient of . Visual inspection revealed more instances where some clinicians annotated areas that others did not, indicating greater uncertainty in depth judgment. Soft comparison results suggest that even when clinicians agree on the depth of a burn area, they may still disagree on its precise boundaries. These findings are consistent with previous studies concluding significant disagreement in visual burn depth assessment by clinicians, but our work reveals the specific manifestations and extent of this disagreement at the pixel level for the first time. Considering the substantial disagreement observed among human clinicians, we recognize that visual assessment of burn wounds and precise annotation present significant challenges. This inherent variability in gold-standard labels inevitably constrains the achievable accuracy of deep learning models, as the training targets themselves contain considerable uncertainty. The performance ceiling of any supervised learning approach is fundamentally bounded by the quality and consistency of the annotated data used for training. Compared to U-Net and DeepLabV3Plus, SAM-DR achieves better performance through simple architectural changes. It also preserves useful pre-trained knowledge from large-scale training. This demonstrates the value of transferring such knowledge to medical image analysis, which aligns with findings from MedSAM. The depth regression approach represents a promising alternative method worth further exploration. The introduction of interactive threshold adjustment effectively balances automation and expert experience. Compared to fixed thresholds, manual threshold adjustment brings significant performance improvements across all models, especially for region 3# burns, where SAM-DR’s Dice coefficient improves by 0.685. This substantial improvement primarily stems from the manual thresholding process effectively suppressing false positives and completely false negative samples during the selection of optimal thresholds. This improvement not only reflects the importance of algorithmic optimization but, more importantly, provides an effective mechanism for integrating clinical expertise into AI output. However, it must be emphasized that in clinical diagnostic scenarios, our approach provides continuous depth estimation rather than thresholded results, serving as reference information only and not as a direct diagnostic tool. The tool we developed successfully translates research findings into practical applications, demonstrating value in both clinical assessment and data annotation scenarios. The tool’s multi-mode operational design allows it to adapt to various clinical needs, providing effective support from detailed regional analysis to rapid screening. Seamless integration with existing annotation workflows further reduces the barrier to use, providing technical support for building large-scale, high-quality datasets. This study has several limitations worthy of attention. First, the accurate diagnosis of full-thickness burns is critically important in clinical practice, as it often directly determines surgical intervention decisions. However, the limited number of full-thickness burn samples in our dataset constrains the model’s validated performance on this essential category, reflecting the practical challenge of collecting rare but severe burn cases. Second, our data solely originate from a pediatric population, and generalizability to adult burn assessment requires further validation. Furthermore, all data were collected during debridement approximately 48 hours post-burn, which, while aligning with the clinically accepted optimal assessment timing, limits the model’s applicability at other time points. Additionally, our dataset exclusively comprises individuals of Asian ethnicity, leaving the model’s generalizability to other racial groups unknown. Conclusion This study introduces SAM-DR, an innovative framework that redefines burn depth assessment by transitioning from traditional discrete classification to continuous depth regression. Through meticulous architectural modifications to the Segment Anything Model, our approach achieves exceptional burn wound segmentation performance with a Dice score of 0.96, while simultaneously providing clinically relevant depth estimation that surpasses human consensus in accuracy. The comprehensive pixel-level analysis of inter-rater variability provides unprecedented insights into the substantial subjectivity inherent in visual burn assessment, particularly evident in the evaluation of deeper burn regions where diagnostic consensus remains challenging. The integration of interactive thresholding establishes an effective paradigm for combining algorithmic precision with clinical expertise, enabling flexible adaptation to both diagnostic and research scenarios. Our web-based tool successfully bridges the gap between technological innovation and clinical workflow, delivering both objective diagnostic support and efficient data annotation capabilities in a unified platform. This dual functionality addresses critical needs in contemporary burn care by enhancing diagnostic consistency while facilitating the creation of large-scale, high-quality datasets for future research. While limitations exist regarding sample size diversity and population specificity, our work establishes a robust foundation for completely non-contact, accessible burn assessment with particular relevance for resource-limited settings. The methodological framework of adapting foundational models through targeted architectural changes, complemented by interactive refinement mechanisms, presents a transferrable approach that may find meaningful applications across various medical imaging domains requiring precise tissue characterization. Future efforts will focus on expanding dataset diversity–through both continued clinical collection and the exploration of synthetic data generation techniques–enhancing model generalizability across demographic populations, and conducting rigorous clinical validation to translate this technological advancement into tangible improvements in patient care outcomes. Acknowledgements We thank the clinical staff of Shenzhen Children’s Hospital for their assistance with data collection and coordination. We also acknowledge the technical support provided by the research and engineering team. Author contributions H.W. conceived the study, designed the methodology, conducted all experiments, performed the data analysis, and wrote the manuscript. S.Z., W.L., X.C., Q.M., H.C., L.F., Z.Z, T.S. and Z.X. contributed to data collection, clinical assessment, and annotation. S.T. and K.Z. provided technical guidance, supervised model development, and revised the manuscript. K.Z., Y.L., and Z.X. supervised the overall project, provided critical revisions, and approved the final manuscript as corresponding authors. All authors reviewed and approved the final version of the manuscript. Funding This study was supported by grants from the Guangdong Province Graduate Education Innovation Project (2025JGXM_149). Data availability The datasets generated and analyzed during the current study are not publicly available due to patient privacy concerns but are available from the corresponding author on reasonable request and with permission from the institutional ethics committee. 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. Contributor Information Kaize Zheng, Email: kzzheng0130@163.com. Yanyan Liang, Email: yyliang@must.edu.mo. Zhu Xiong, Email: bamboobear@163.com. References 1. Lawrence, J. W., Mason, S. T., Schomer, K. & Klein, M. B. Epidemiology and impact of scarring after burn injury: a systematic review of the literature. J. Burn Care Res. 33 , 136–146 (2012). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Meng, F. et al. Pediatric burn contractures in low-and lower middle-income countries: A systematic review of causes and factors affecting outcome. Burns 46 , 993–1004 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 3. Phelan, H. A. et al. Use of 816 consecutive burn wound biopsies to inform a histologic algorithm for burn depth categorization. J. Burn Care Res. 42 , 1162–1167 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Shin, J. Y. & Yi, H. S. Diagnostic accuracy of laser doppler imaging in burn depth assessment: Systematic review and meta-analysis. Burns 42 , 1369–1376 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 5. Taib, B. G. et al. Artificial intelligence in the management and treatment of burns: A systematic review and meta-analyses. J. Plastic Reconstruct. Aesthetic Surg. 77 , 133–161 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 6. Zhang, R., Tian, D., Xu, D., Qian, W. & Yao, Y. A survey of wound image analysis using deep learning: classification, detection, and segmentation. IEEE Access 10 , 79502–79515 (2022). [ Google Scholar ] 7. Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention , 234–241. (Springer, 2015). 8. Yadav, D. et al. Spatial attention-based residual network for human burn identification and classification. Sci. Rep. 13 , 12516 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Zhang, X. et al. Multi-feature extraction and selection method to diagnose burn depth from burn images. Electronics 13 , 3665 (2024). [ Google Scholar ] 10. Ethier, O. et al. Using computer vision and artificial intelligence to track the healing of severe burns. J. Burn Care Res. 45 , 700–708 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Rahman, M. M. et al. A framework for advancing burn assessment with artificial intelligence. Military Med. 190 , 387–393 (2025a). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Rahman, M. M. et al. Ai-driven integrated system for burn depth prediction with electronic medical records: Algorithm development and validation. JMIR Med. Inform. 13 , e68366 (2025b). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Bulut, C., Kolca, D. & Tarlak, F. Development a software for detecting burn severity using convolutional neural network-based approach. Sigma 43 , 598–606 (2025). [ Google Scholar ] 14. Yıldız, M. et al. Segmentation and classification of skin burn images with artificial intelligence: Development of a mobile application. Burns 50 , 966–979 (2024). [ DOI ] [ PubMed ] 15. Li, Z. et al. Gl-fusionnet: fusing global and local features to classify deep and superficial partial thickness burn. Math. Biosci. Eng. 20 , 10153–10173 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 16. Pabitha, C. & Vanathi, B. Dense mesh rcnn: assessment of human skin burn and burn depth severity. J. Supercomput. 80 , 1331–1362 (2024). [ Google Scholar ] 17. Li, X., Liu, Z. & Liu, L. Pediatric burnnet: Robust multi-class segmentation and severity recognition under real-world imaging conditions. SAGE Open Med. 13 , 20503121251360090 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Abdolahnejad, M. et al. Novel cnn-based approach for burn severity assessment and fine-grained boundary segmentation in burn images. IEEE Trans. Instrument. Measure. (2025). 19. Tan, P., Nyeko-Lacek, M., Walsh, K., Sheikh, Z. & Lewis, C. J. Artificial intelligence-enhanced multispectral imaging for burn wound assessment: Insights from a multi-centre uk evaluation. Burns 107550 (2025). [ DOI ] [ PubMed ] 20. Kirillov, A. et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision , 4015–4026 (2023). 21. Ma, J. et al. Segment anything in medical images. Nat. Commun. 15 , 654 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Liu, H., Yue, K., Cheng, S., Li, W. & Fu, Z. A framework for automatic burn image segmentation and burn depth diagnosis using deep learning. Comput. Math. Methods Med. 2021 , 5514224 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Xie, J., Li, H., Li, J. & Xu, X. Ltb-net: A lightweight transformer-based burn depth segmentation network. In BIBE 2024; The 7th International Conference on Biological Information and Biomedical Engineering , 234–239 (2024). 24. Zhang, D. & Xie, J. Semi-supervised burn depth segmentation network with contrast learning and uncertainty correction. Sensors 25 , 1059 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Chang, C. W. et al. Application of multiple deep learning models for automatic burn wound assessment. Burns 49 , 1039–1051 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 26. Rozo, A. et al. A deep learning image-to-image translation approach for a more accessible estimator of the healing time of burns. IEEE Trans. Biomed. Eng. 70 , 2886–2894 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 27. Ji, S., Xiao, S., Xia, Z., of Burns, C. B. A. T. R. & Trauma Committee, C.-S. M. E. A. o. C. Consensus on the treatment of second-degree burn wounds (2024 edition). Burns Trauma 12 , tkad061 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 28. Wang, W. Advanced auto labeling solution with added features. https://github.com/CVHub520/X-AnyLabeling (2023). 29. Cubuk, E. D., Zoph, B., Shlens, J. & Le, Q. V. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , 702–703 (2020). 30. Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV) , 801–818 (2018). 31. Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J. & Zisserman, A. The PASCAL visual object classes challenge 2012 (VOC2012) results. http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html . Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The datasets generated and analyzed during the current study are not publicly available due to patient privacy concerns but are available from the corresponding author on reasonable request and with permission from the institutional ethics committee. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.7 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

• Pulmonary Arterial Hypertension and Associated Cardiovascular Disease Detection Using Artificial Intelligence#858693
• CARDIOLOGIST Trial: Artificial Intelligence Enabled Electrocardiogram for Atrial Fibrillation Detection#1103054
• Detection of Skin Cancer Using Deep Learning and Image Processing#733632
• Evaluation of Speech and Voice Biomarkers in Association With Burnout#900739
• Burns v. South Muskoka Memorial Hospital, Board of Inquiry, July 1994 BOI 94-018#334010