Development of a multi-task learning framework with gradnorm for precise wound tissue analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Chronic wounds present significant patient and healthcare challenges.
• Accurate wound analysis, including size and tissue composition, is crucial for effective treatment.
• Existing segmentation methods like Separate Task Learning (STL) are inefficient, while Multi-Task Learning (MTL) can suffer from task imbalance.

Purpose of the Study:

• To introduce WING-MTL, a novel Multi-Task Learning framework for wound and wound tissue segmentation.
• To address the task imbalance and performance degradation issues in conventional MTL approaches.
• To enhance accuracy, training stability, and parameter efficiency in chronic wound analysis.

Main Methods:

• Developed WING-MTL (Wound and Wound Tissue Integrated with Gradient Normalization Multi-Task Learning) framework.
• Utilized an Attention-UNet backbone with real-time Gradient Normalization for balanced optimization.
• Evaluated WING-MTL across diverse architectures (UNet, Resnet, Transformer) and performed longitudinal patient analysis.

Main Results:

• WING-MTL demonstrated statistically significant improvements over STL and conventional/advanced MTL methods.
• Achieved balanced learning with both tasks converging at the same epoch.
• Showcased superior segmentation performance, especially for challenging tissues like slough and epithelium.
• Validated consistent performance across various architectures and demonstrated clinical utility in longitudinal studies.

Conclusions:

• WING-MTL effectively balances gradient magnitudes across tasks, enhancing accuracy and stability in wound segmentation.
• The framework maintains parameter efficiency while overcoming MTL task imbalance issues.
• WING-MTL offers a promising, accurate, and versatile approach for tracking wound healing and supporting clinical decisions.