🤖 AI

Improving Automated Wound Assessment Using Joint Boundary Segmentation and Multi-Class Classification Models

This study introduces a YOLOv11-based deep learning model that simultaneously performs wound boundary segmentation and multi-class classification across five wound types, demonstrating that data augmentation significantly enhances performance and that the lightweight YOLOv11n variant offers a robust, resource-efficient solution for real-time clinical and remote wound assessment.

Original authors: Mehedi Hasan Tusar, Fateme Fayyazbakhsh, Igor Melnychuk, Ming C. Leu

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mehedi Hasan Tusar, Fateme Fayyazbakhsh, Igor Melnychuk, Ming C. Leu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to heal a patient's wound. In the old days, you'd have to squint at the injury, guess how big it is with a ruler, and try to figure out exactly what kind of wound it is (a burn, a pressure sore, a diabetic ulcer, etc.). It's like trying to identify a specific type of cloud in a stormy sky while holding a flashlight that flickers. It's slow, it's subjective, and different doctors might see different things.

This paper introduces a super-smart digital assistant (an AI) that acts like a "super-eye" for doctors. It doesn't just look at the wound; it does two jobs at once, instantly and with incredible precision.

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: The "One-Tool" Limitation

Before this study, most AI tools were like Swiss Army knives that only had one blade. Some could only draw a line around the wound (segmentation), and others could only guess the type of wound (classification). They couldn't do both at the same time, and they often only knew how to handle one or two types of wounds, ignoring the rest.

2. The Solution: The "All-in-One" Detective

The researchers built a new AI model based on YOLOv11 (which stands for "You Only Look Once"). Think of this model as a highly trained detective who walks into a room and immediately:

Draws a perfect outline around the wound (telling you exactly how big it is).
Identifies the culprit (telling you if it's a Burn, a Pressure Injury, a Diabetic Ulcer, a Vascular Ulcer, or a Surgical Wound).

They taught this detective using a massive library of 2,963 photos of real wounds. To make sure the detective didn't just memorize the photos but actually learned the rules, they used a technique called Cross-Validation. Imagine giving the detective five different practice exams, rotating the questions so they see every type of wound in every possible combination. This ensures they are truly smart, not just lucky.

3. The "Data Augmentation" Trick: The Chameleon Effect

One big problem with wound photos is that lighting changes, skin tones vary, and sometimes the wound is hard to see (like a faint burn on pale skin). The AI struggled a bit with these "faint" cases.

To fix this, the researchers used Data Augmentation. Imagine you are teaching a child to recognize a dog. You show them a photo of a dog. Then, you show them the same photo but:

Turned sideways (rotation).
Flipped upside down (flipping).
Made brighter or darker (lighting changes).
Made slightly blurry or colorful.

By doing this, the AI learned that a wound is a wound, even if the lighting is weird or the angle is strange. It became a chameleon, able to spot wounds in any condition. This trick made the AI significantly better at spotting the tricky, low-contrast burns.

4. The "Size" Choice: The Race Car vs. The Hybrid

The researchers tested five different versions of this AI, ranging from a tiny, lightweight version to a massive, heavy-duty version.

The Heavyweight (YOLOv11x): This is like a Formula 1 race car. It is the fastest and most accurate, but it requires a lot of fuel (computer power) and is expensive to run. It got the highest scores.
The Lightweight (YOLOv11n): This is like a fuel-efficient hybrid car. It's much smaller and faster to train, using only 25% of the time and resources of the big one. While it's slightly less "perfect," it's still incredibly accurate.

Why does this matter? The "Hybrid" version is perfect for a doctor's tablet or a mobile phone in a remote village. You don't need a supercomputer to get a great diagnosis; you just need a smart phone app.

5. The Results: Beating the Old Guard

When they compared their new AI to older, famous models (like VGG16 or ResNet), the new YOLOv11 model won the race.

Accuracy: It correctly identified the wound type and drew the boundary with high precision.
Versatility: It handled all five types of wounds, whereas older models often got confused or ignored certain types.
Confidence: The AI was so sure of its answers that if you asked it to be 96% sure, it was still right almost every time.

The Bottom Line

This paper is about giving doctors a magic magnifying glass that never gets tired, never gets distracted, and sees things the human eye might miss.

For the Doctor: It saves time and reduces guesswork.
For the Patient: It means faster, more accurate treatment plans.
For the Future: Because the "lightweight" version works so well, we can soon have this technology on a smartphone, allowing people in remote areas to get expert-level wound care without needing to travel to a big hospital.

In short, they took a complex medical problem and solved it with a digital tool that is both powerful enough for a hospital and small enough for a pocket.

1. Problem Statement

Chronic and acute wound management faces significant challenges due to the subjectivity, time consumption, and inter-observer variability of manual clinical assessments. While Artificial Intelligence (AI) and Deep Learning (DL) offer automation solutions, existing models suffer from several critical limitations:

Task Isolation: Most models perform either wound boundary segmentation (WBS) or wound classification (WC) as separate tasks, rather than jointly.
Limited Scope: Many models are trained on narrow datasets focusing on only one or two wound types (e.g., only Diabetic Foot Ulcers), neglecting others like Burn Injuries (BI), Pressure Injuries (PI), Vascular Ulcers (VU), and Surgical Wounds (SW).
Data Scarcity & Imbalance: High-performance models often rely on small, imbalanced datasets, leading to poor generalization, especially for visually subtle or low-contrast wounds like early-stage burns.
Computational Constraints: Many state-of-the-art models are too computationally heavy for deployment in resource-constrained Point-of-Care (POC) or mobile settings.

2. Methodology

The study proposes a unified deep learning framework based on YOLOv11 (released by Ultralytics in October 2024) to simultaneously perform instance segmentation and multi-class classification.

Dataset Preparation

Source: 2,963 high-quality, de-identified wound images were collected from public repositories (GitHub, Roboflow, Kaggle) and validated by clinicians.
Classes: Five distinct wound categories: Burn Injury (BI), Pressure Injury (PI), Diabetic Foot Ulcer (DFU), Vascular Ulcer (VU), and Surgical Wound (SW).
Augmentation: To address class imbalance and improve generalization (particularly for BI), the dataset was augmented using rotation, flipping, and variations in brightness, saturation, and exposure. This expanded the training set from 2,963 to 7,147 images.
Validation: A stratified five-fold cross-validation approach was used to ensure robust, unbiased evaluation across different data splits.

Model Architecture

The study utilizes YOLOv11, a state-of-the-art architecture featuring:

Backbone: Modified Cross Stage Partial blocks with kernel size 2 (C3k2), replacing the C2f blocks in YOLOv8. This reduces parameters by ~22% while maintaining feature extraction power.
Neck: Incorporates Spatial Pyramid Pooling - Fast (SPPF) and Path Aggregation Network (PANet) for multi-scale feature fusion.
Head: Dual-branch design with an anchor-free detection head for bounding boxes and a segmentation head using a prototype-based approach for pixel-level masks.
Attention: Integration of Convolutional blocks with Parallel Spatial Attention (C2PSA) to focus on significant regions.
Loss Function: A compound loss combining Binary Cross-Entropy (BCE), Intersection over Union (IoU), Cross-Entropy, Distribution Focal Loss (DFL), and Task-Aligned Assignor.

Experimental Setup

Variants Tested: Five YOLOv11 variants were fine-tuned: Nano (n), Small (s), Medium (m), Large (l), and Extra-Large (x).
Tasks:
1. Wound Boundary Segmentation (WBS): Pixel-level instance segmentation.
2. Wound Classification (WC): Multi-class classification of the five wound types.

3. Key Contributions

Unified Multi-Task Framework: First known work to simultaneously perform WBS and WC across five major wound types using a single image input.
Comprehensive Dataset: Creation of a balanced, expert-annotated dataset covering BI, PI, DFU, VU, and SW, with extensive augmentation strategies to handle visual variability.
YOLOv11 Adaptation: Successful fine-tuning of the latest YOLOv11 architecture for medical wound analysis, demonstrating its superiority over previous YOLO versions and other DL models.
Scalability Analysis: A comparative evaluation of model variants (n to x), identifying optimal trade-offs between accuracy and computational cost for different deployment scenarios (e.g., cloud vs. mobile).
Robustness Validation: Extensive benchmarking against human-level performance and existing models (VGG16, ResNet, Mask R-CNN) using confusion matrices, ROC curves, and visual inference.

4. Results

The models were evaluated using Precision, Recall, F1-score, mAP50–95, and AUROC.

Performance on Augmented Dataset

Best Overall Performance: YOLOv11x achieved the highest metrics:
- WBS: F1-score of 0.9341, mAP50 of 0.9629.
- WC: F1-score of 0.8736, mAP50 of 0.9194.
Lightweight Efficiency: YOLOv11n (Nano) provided competitive accuracy (WBS F1: 0.9285; WC F1: 0.8607) with significantly lower computational cost (training time was 25% of YOLOv11x), making it ideal for mobile/POC devices.
Impact of Augmentation: Data augmentation significantly improved performance, particularly for Burn Injuries (BI), which previously struggled due to low contrast.
- YOLOv11n WBS: Precision improved from 0.8936 to 0.9263; Recall from 0.8478 to 0.9307.
- YOLOv11n WC: Precision improved from 0.8701 to 0.8808.

Comparative Analysis

vs. Traditional Classifiers: The YOLOv11 model outperformed VGG16, AlexNet, and ResNet101 in F1-scores (0.8767 vs. 0.8732, 0.8401, and 0.8020, respectively) while also providing segmentation masks.
vs. Segmentation Models: It surpassed DeepLabV3, UNet, and Mask R-CNN in F1-scores and provided faster inference.
Class-Specific Performance:
- High Accuracy: Surgical Wounds (SW) and Pressure Injuries (PI) showed the highest classification accuracy due to distinct morphological features.
- Challenges: Burn Injuries (BI) and Vascular Ulcers (VU) were more difficult to classify due to visual similarity with normal skin and low contrast, though augmentation mitigated this.
- AUC Scores: All classes achieved AUC scores $\ge$ 0.96, with SW reaching a perfect 1.00.

5. Significance and Conclusion

This study demonstrates that YOLOv11 is a highly effective architecture for real-time, automated wound assessment. By unifying segmentation and classification across five diverse wound types, the proposed system addresses the fragmentation in current AI wound care tools.

Clinical Impact: The system offers a standardized, reproducible method for wound monitoring, reducing subjectivity and enabling earlier interventions.
Deployment Viability: The success of the lightweight YOLOv11n variant suggests that high-accuracy wound analysis can be deployed on resource-constrained mobile devices, facilitating remote care and point-of-care diagnostics.
Future Directions: The authors note limitations regarding the need for more diverse skin tones, the integration of clinical metadata (patient history, location), and the optimization of models for embedded systems. Future work will also explore generative AI for data augmentation and semi-supervised learning to reduce annotation burdens.

In summary, Tusar et al. (2025) provide a robust, scalable, and clinically relevant solution that advances the state-of-the-art in automated wound care, bridging the gap between high-performance AI and practical clinical application.