WoundAmbit: Bridging State-of-the-Art Semantic Segmentation and Real-World Wound Care

arXiv:2504.06185 · cs.CV, cs.AI · Submitted 2025-04-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "WoundAmbit: Bridging State-of-the-Art Semantic Segmentation and Real-World Wound Care".

Tom: The gist: Automated wound monitoring via mobile image capture can reduce in-person physician visits by enabling remote tracking of wound size through semantic segmentation.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So what they actually did in this "WoundAmbit" paper is build this comprehensive comparison by taking twelve state-of-the-art deep learning models from general vision, medical imaging, and specific wound challenges.

Jane: They make sure everything is standardized: the training settings, the way they augment the data, and even the input tensor dimensions so every model gets an equal chance to compete <ref:2504.06185#pg8>.

Lu: They organized their selection into groups like DFU segmentation, medical segmentation, and general GP models to make sure they had a good mix of architectures—things like CNNs, ViTs, and hybrid methods—all in one study <ref:2504.06185#pg4>.

Meng: For the general purpose models specifically, they focused on architectures that had around fifty to sixty million parameters because they found a good balance between performance and being able to run on limited hardware <ref:2504.06185#pg9>.

Lalam: Plus, they used a specific dataset of three hundred forty-three wound images taken at different body sites to test these models against real variety <ref:2504.06185#pg8>.

The paper's summary: Tom: The authors suggest a few ways to make this system better, and the most important thing they propose is using this reference object approach to turn the AI’s segmented mask into an actual size measurement.

Jane: This reference object approach uses four ArUco markers that are twelve millimeters on each side, plus a four millimeter margin of white space, which helps accurately detect the wound size independently of just the AI segmentation <ref:2504.06185#pg8>.

Lu: They use multiple ArUco markers whenever possible and calculate ratios using predefined real-world distances between marker pairs to reduce perspective distortion and measurement noise <ref:2504.06185#pg8>.

Meng: This means the size estimation process doesn't depend on training a separate model just to find the reference object, which simplifies the overall system architecture <ref:2504.06185#pg3>.

Lalam: They also focus on interpretability by using Grad-CAM visualizations to show exactly where the AI is looking when it makes its decision, which helps build trust between a clinician and the tool <ref:2504.06185#pg3>.

The paper's improvements: Tom: So wrapping up this study on "WoundAmbit," it shows that transformer-based models like TransNeXt did really well in terms of generalizability when tested on out of distribution data <ref:2504.06185#pg10>.

Jane: Even with some variations in inference speed, all the models managed to process at least one image per second on a CPU, which is pretty adequate for what they are trying to achieve <ref:2504.06185#pg10>.

Lu: The findings also point out that the top five models on their out of distribution dataset were general purpose architectures, which suggests these modern vision and segmentation models are stronger than specialized medical techniques when dealing with new data <ref:2504.06185#pg10>.

Meng: They found that while no single architecture was clearly the best for size retrieval, VW-Conv was favored by experts for its mask selection quality, and the measurement errors were quite low—height estimates were under four point four millimeters and width under three point two millimeters <ref:2504.06185#pg10>.

Lalam: The overall framework, WoundAmbit, is now being put into a custom telehealth system so that when doctors need it, they get a detailed view of the wound area right there <ref:2504.06185#pg10>.

Tom: To sum up this work on "WoundAmbit," this study advances automated wound segmentation by testing twelve state-of-the-art deep learning methods from general vision and medical imaging against rigorous standards on both known and unknown data <ref:2504.06185#pg1>.

Jane: It’s a step toward bridging the gap between complex AI techniques and practical, remote wound care monitoring by creating a robust framework for size estimation <ref:2504.06185#pg2>.

Lu: Future work they mention is expanding the dataset with smartphone images from more diverse patient demographics and conditions to further fine-tune these models <ref:2504.06185#pg10>.

Meng: And integrating this estimation framework with existing clinical systems, along with assessing patient-reported outcomes in clinical trials, is what’s next for real-world impact <ref:2504.06185#pg10>.

Lalam: This research represents an important step toward improving remote wound size monitoring and enhancing the quality of patient care through this AI-driven approach to wound measurement <ref:2504.06185#pg1>.

Conclusion: Tom: So we just went through "WoundAmbit: Bridging State-of-the-Art Semantic Segmentation and Real-World Wound Care," where they really connect those advanced segmentation models to actual wound care on a phone.

Jane: They tackled that problem of manual measurements being inaccurate because of irregular wound shapes or just how subjective it is to measure them by hand <ref:2504.06185#pg2>.

Lu: They did this by benchmarking twelve state-of-the-art models against each other under standardized conditions using five fold cross validation to keep the comparison fair <ref:2504.06185#pg7>.

Meng: And they specifically looked at how well these models generalize to out of distribution data, which is super important because a model that works great on one wound type might totally fail on a new one <ref:2504.06185#pg3>.

Lalam: Plus, they focused on efficiency too, because for this application to actually work on a mobile device, the AI needs to run fast and use not too many resources <ref:2504.06185#pg10>.

Tom: What they did in that benchmark was really comprehensive. They selected twelve models from wound-specific studies, medical imaging, and general purpose vision models to make sure they covered all the bases <ref:2504.06185#pg4>.

Jane: They standardized everything then: the training settings, the data augmentation pipeline, and even the input tensor dimensions so every model got a fair shot at comparison <ref:2504.06185#pg8>.

Lu: They organized their selection into groups like DFU segmentation, medical segmentation, and general GP models to ensure they had architectural diversity—things like CNNs and ViTs all in one study <ref:2504.06185#pg4>.

Meng: For those general purpose models specifically, they focused on architectures with around sixty million parameters because that was where they found a good trade-off between performance and being computationally feasible <ref:2504.06185#pg9>.

Lalam: They also used a dedicated dataset of three hundred forty-three wound images taken at different body sites to test these models against real variety <ref:2504.06185#pg8>.

Tom: So they didn't stop there with just the segmentation part. The authors suggest a few key improvements, and the biggest one is introducing this reference object approach to turn the AI’s segmented mask into an actual size measurement.

Jane: This reference object approach uses four ArUco markers with twelve millimeter sides and a four millimeter margin of white space to help accurately detect the wound size, making it independent of just the AI segmentation itself <ref:2504.06185#pg8>.

Lu: They use multiple ArUco markers whenever possible and compute ratios based on predefined real-world distances between marker pairs to reduce perspective distortion and measurement noise <ref:2504.06185#pg8>.

Meng: This means the size estimation process doesn't depend on training a separate model just to find the reference object, which simplifies the overall system architecture <ref:2504.06185#pg3>.

Lalam: They also focus on interpretability with Grad-CAM visualizations to show exactly where the AI is looking when it makes its decision, which helps build trust between a clinician and the tool <ref:2504.06185#pg3>.

Tom: So wrapping up this discussion on "WoundAmbit," the study shows that transformer-based models like TransNeXt performed very well in terms of generalizability on out of distribution data <ref:2504.06185#pg10>.

Jane: Despite some variations in how fast they ran, all the models were able to process at least one image per second on a CPU, which is pretty adequate for what they're aiming to do <ref:2504.06185#pg10>.

Lu: The findings also show that the top five models on their out of distribution dataset were general purpose architectures, which suggests these modern vision and segmentation models are stronger than specialized medical techniques when dealing with new data <ref:2504.06185#pg10>.

Lalam: The overall framework, WoundAmbit, is now being integrated into a custom telehealth system to give physicians a detailed view of the wound area when they need it <ref:2504.06185#pg10>.

Jane: It’s a step toward bridging the gap between complex AI techniques and practical, remote wound care monitoring by creating a robust framework for size estimation <ref:2504.06185#pg2>.

Lu: Future work they mention is expanding the dataset with smartphone images from more diverse patient demographics and conditions to further fine-tune these models <ref:2504.06185#pg10>.

Julius-Maximilians-University Würzburg University Hospital Würzburg

cs.CV, cs.AI

Submitted: 2025-04-08

Updated: 2026-10-09

Comments: Main paper: 18 pages; supplementary material: 15 pages

Journal ref: Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track (ECML PKDD 2025). Lecture Notes in Computer Science, vol. 16021, pp. 285-303. Springer, Cham (2026)

DOI: 10.1007/978-3-032-06118-8_17

Code: https://github.com/MrYxJ/calculate-flops.pytorch

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The gist: Automated wound monitoring via mobile image capture can reduce in-person physician visits by enabling remote tracking of wound size through semantic segmentation.

Key concepts

Semantic Segmentation: This is a type of deep learning where an AI classifies every single pixel in an image as belonging to a specific category, such as 'wound' or 'skin.' It creates a precise digital mask that outlines the exact shape and boundary of the wound, which is crucial for accurately measuring its size.
Reference Object (RO)-based Approach: To convert the AI's pixel mask into a real-world measurement, this method uses physical markers—ArUco markers—placed in the image. By detecting these known objects and calculating the ratio between them, the system can accurately estimate the actual size of the segmented wound in millimeters.
Generalizability (OOD Data): This refers to how well a model performs on new, unseen data that differs from its training set. The study tested models on an Out-of-Distribution (OOD) dataset to ensure they can accurately segment wounds taken in different settings or conditions, indicating a robust and reliable AI tool for real-world use.

Terminology

Interpretability — 2× in this paper · explained in 15 episodes
Accuracy — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Data Augmentation — 1× in this paper · explained in 7 episodes
Generalizability — 1× in this paper · explained in 6 episodes
Generalization — 1× in this paper · explained in 13 episodes
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist: Automated wound monitoring via mobile image capture can reduce in-person physician visits by enabling remote tracking of wound size through semantic segmentation.

How it works

Comprehensive DL Benchmark: The study conducts a rigorous benchmarking study by systematically selecting 12 SotA DL architectures from wound-specific, medical, and GP vision models, covering a diverse range of design paradigms including CNNs, ViTs, and hybrid models<ref:2504.06185#pg2> To ensure comparability, all models are trained under standardized conditions on publicly available data and evaluated using 5-fold cross-validation (CV)<ref:2504.06185#pg7> The evaluation framework emphasizes clinically relevant properties by assessing generalizability on a dedicated OOD dataset<ref:2504.06185#pg8> Specifically created for this work, the dataset comprises 343 wound images taken at various body sites<ref:2504.06185#pg8> Moreover, we analyze computational efficiency, including trainable parameters, Giga Multiply-Accumulate Operations (GMACs), and inference times on both GPU and CPU<ref:2504.06185#pg10> Lastly, we investigate model interpretability using Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations to assess clinically relevant decision-making.

Model Selection

The selection process for the 12 architectures was guided by several criteria<ref:2504.06185#pg4> The categories included:

(I) DFU segmentation:

(II) Medical segmentation:

(III) GP SS models:

The final selection aimed for a balance between domain specificity, cross-domain variety, and methodological diversity<ref:2504.06185#pg5> The criteria prioritized architectural diversity to ensure a comprehensive representation of various design paradigms including CNN-based, ViT-based, and hybrid methods<ref:2504.06185#pg8> Computational efficiency was also a key factor, prioritizing models that offer a good trade-off between performance and computational feasibility<ref:2504.06185#pg8> For GP models, the selection favored architectures with approximately 50–60M parameters<ref:2504.06185#pg9>

Unified Training Procedure and Evaluation

A unified benchmarking environment was established by standardizing the training process All deep learning methods were implemented with their architecture-specific settings, using ImageNet-pretrained backbones for all methods Training settings were standardized across all models, including input tensor dimensions, the number of training epochs, early stopping criteria, optimizer type, loss function, and the data augmentation pipeline A preliminary hyperparameter tuning step was conducted for each architecture to optimize the learning rate within a predefined search range The standardized evaluation procedure employed metrics such as mean Intersection over Union (mIoU), Dice Similarity Coefficient (mDSC), precision (mPrc), and recall (mRec) In terms of model efficiency, the report included trainable parameters and GMACs, along with mean inference time for GPU and CPU execution Finally, explainability was assessed using Grad-CAM-based visualizations.

Real-World Deployment and Size Retrieval

To translate AI-generated wound masks into clinically meaningful size estimates, a reference object (RO)-based approach was developed This RO features four ArUco markers, each with a 12 mm side length and a 4 mm margin of white space to facilitate accurate detection The size retrieval process involves detecting the ArUco markers using OpenCV’s ArucoDetector A robust approach utilizes multiple ArUco markers whenever possible, computing the ratio using predefined real-world distances between specific marker pairs to reduce perspective distortion and measurement noise The final size estimation involves converting segmented wound areas into square millimeters using this ratio and determining the wound’s height and width following standard clinical measurement practices. Evaluation involved collecting N = 20 diverse wound images using the RO and an iPhone 11 Size retrieval quality was quantified using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) Expert assessments involved three dermatologists reviewing AI-generated masks, with metrics defined as Clinical Mask Approval (CMA) and Expert Choice Rate (ECR).

Key Findings

The study demonstrated that the transformer-based TransNeXt showed the highest levels of generalizability<ref:2504.06185#pg10> Despite variations in inference times, all models processed at least one image per second on the CPU, which is deemed adequate for the intended application<ref:2504.06185#pg10> Interpretability analysis typically revealed prominent activations in wound regions, emphasizing focus on clinically relevant features Expert evaluation showed high mask approval for all analyzed models, with VWFormer and ConvNeXtS backbone performing the best<ref:2504.06185#pg10> Size retrieval accuracy was similar across models, and predictions closely matched expert annotations<ref:2504.06185#pg10>. Overall, the AI-driven wound size estimation framework, WoundAmbit, is integrated into a custom telehealth system. The top five models on the OOD dataset were GP architectures, further emphasizing the superiority of contemporary SS and vision models over specialized medical techniques in handling OOD data<ref:2504.06185#pg10>. While no architecture clearly outperformed the others in size retrieval, VW-Conv was most favored by experts for best mask selection. The CMA ranged from 83.3% to 86.7% across models, reflecting high mask quality and strong physician approval. On average, height estimates were slightly overestimated, while width was slightly underestimated. Consistently low MAE (≤4.4 mm height, ≤3.2 mm width) and MAPE (≤14.1% height, ≤16.2% width) indicate robust size retrieval across wound sizes. Future work will focus on expanding the dataset with smartphone images from diverse patient demographics and conditions for model finetuning. The ongoing longitudinal study aims to empirically assess the effectiveness of the overall telehealth solution, with a focus on patient adherence and wound care outcomes.

Conclusion

This work advances automated wound segmentation by benchmarking 12 SotA DL methods from GP vision, medical segmentation, and wound care Our evaluation on both in-distribution and OOD datasets demonstrates that modern GP models, especially TransNeXt and VWFormer, exhibit strong generalization and remain computationally feasible for clinical use While our AI-driven wound size estimation framework, WoundAmbit, shows strong clinical promise, future work will focus on expanding the dataset with smartphone images from diverse patient demographics and conditions for model finetuning. Additionally, integrating WoundAmbit with existing clinical systems and assessing patient-reported outcomes in clinical trials will be crucial for evaluating its real-world impact. Lastly, incorporating the OOD dataset and prospectively collected smartphone images into the training is expected to improve model performance further. Overall, this work represents an important step toward bridging recent advances in computer vision with real-world wound care, offering the potential to improve remote wound size monitoring and enhance patient care. The code and supplementary material are available on GitHub and Zenodo, respectively. The paper is structured as follows: Section 2 describes the DL benchmark (Fig. 1, left), covering model selection, methodology, datasets, and results Section 3 details the size retrieval process and its evaluation, along with the integration of WoundAmbit into a specially designed telehealth system (Fig.

Improvements for AI systems

What is an improved AI system can perform automated wound size estimation from RGB images in a mobile setting by integrating a Reference Object (RO)-based approach to convert AI-generated masks into clinically relevant wound size estimates. This overcomes the limitation of existing methods that do not address the conversion of segmentation masks into real-world wound size.
What is an improved AI system can provide a comprehensive DL benchmark by systematically selecting 12 SotA DL architectures from wound-specific, medical, and GP vision models and evaluating them using 5-fold cross-validation (CV) to ensure comparability across design paradigms. This rigorous benchmarking addresses the limitation of limited adoption of SotA Vision Models for wound analysis.
What is an improved AI system can offer real-world deployment capabilities by developing a reference object (RO)-based approach that leverages ArUco marker detection, making it independent of the AI algorithm used for segmentation. This ensures precise object size estimation independent of AI algorithms, eliminating the need for including RO detection as an additional class in model training.
What is an improved AI system can improve its robustness by assessing performance on out-of-distribution (OOD) data, as demonstrated by the finding that TransNeXt demonstrated the strongest generalization and models maintained competitive performance relative to CFU data on OOD images with minimal drop in mIoU.
What is an improved AI system can enhance its clinical relevance by providing interpretability insights through Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations which reveal prominent activations in wound regions, emphasizing focus on clinically relevant features. This allows for a deeper understanding of model behavior and fosters greater trust among clinicians.
What is an improved AI system can be made computationally efficient by assessing metrics such as trainable parameters, Giga Multiply-Accumulate Operations (GMACs), and inference times to identify trade-offs between efficiency and performance, noting that models such as U-Net, MISSFormer, and HiFormer excel in inference speed on both GPU and CPU.
What is an improved AI system can be integrated into a practical telehealth solution by embedding the estimation framework into a custom telehealth platform with a dedicated microservice for AI processing to provide physicians with a detailed wound view displaying the AI-predicted wound area (○1).