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

## TL;DR

A new multi-task learning framework called WING-MTL improves wound and tissue analysis accuracy by balancing gradients during training, offering better performance than traditional methods.

## Contribution

WING-MTL introduces gradient normalization to address task imbalance in multi-task learning for wound tissue segmentation.

## Key findings

- WING-MTL shows statistically significant improvements over separate task learning and other MTL methods in wound tissue segmentation.
- The framework achieves balanced learning by converging both tasks at the same epoch, improving accuracy for challenging tissue types like slough and epithelium.
- WING-MTL demonstrates architectural flexibility and consistent performance across diverse models like UNet, Resnet, and Transformers.

## Abstract

Chronic wounds impose a substantial burden on patients and healthcare systems, necessitating accurate qualification of precise wound analysis for effective diagnosis and treatment. Wound size and the proportional composition of internal tissues are critical indicators of healing progression. Traditional segmentation approaches such as Separate Task Learning (STL) suffer from parameter inefficiency, while Multi-Task Learning (MTL), though efficient, often experiences task imbalance that leads to performance degradation in specific tasks. To overcome these challenges, this study proposes WING-MTL (Wound and Wound Tissue Integrated with Gradient Normalization Multi-Task Learning), a novel MTL framework that dynamically balances gradient magnitudes across tasks to enhance accuracy and training stability. Implemented on an Attention-UNet backbone,and incorporates Gradient Normalization to adjust learning gradients in real time, ensuring balanced optimization for wound and wound tissue segmentation tasks. Quantitative evaluations demonstrate that WING-MTL yields statistically significant improvements over STL and outperforms conventional and advanced MTL methods. Analysis of validation loss revealed convergence of both tasks at the same epoch, indicating balanced learning. Furthermore, improved segmentation performance was observed compared to both STL and MTL approaches, particularly for challenging wound tissue types such as slough and epithelium in qualitative analysis. These findings confirm that WING-MTL addresses the task imbalance inherent in MTL frameworks while maintaining parameter efficiency. WING-MTL was evaluated across diverse backbones, including not only UNet-based model but also Resnet and Transformer-based architectures to validate the architectural flexibility of the proposed framework. Consistent performance improvement across these varied architectures in wound tissue segmentation demonstrates the broad applicability. Furthermore, longitudinal analysis of chronic wound patients was conducted to assess the clinical utility of WING-MTL in real-world scenarios. The framework presents a promising and accurate approach for tracking wound healing progression and serves as a potential adjunct for clinical decision-making in chronic wound care.