Publication

Abstract : It is challenging to effectively manage wounds that remain open for a longer time. Advanced AI models are effective in the early detection of wounds and diagnosing their underlying causes. However, there are potential biases in these models, such as ethnic, racial, gender, and clinical. Hence, the advancement in healthcare, which should be equally useful to the entire population around the world, fails to achieve this goal, mostly due to ethnic disparity. In this study, a novel database of Indian skin tone general wounds is collected in an uncontrolled setup and mixed with the open AZH dataset to develop a diverse database unbiased to a special ethnic category. Several versions of YOLO algorithms are widely employed in wound detection because of their lightweight design and better performance. Here, a robust and faster frame rate model YOLOv8, is trained with this heterogeneous dataset to reduce the effect of skin tone issues. Using an unrelated Medetec database to evaluate its robustness, the model achieves a precision of 84% and a mean average precision of around 70%. Hence, the proposed model achieves its goal of detecting wounds of varying skin complexions to a considerable extent. The developed unbiased detection model could be integrated into subsequent wound segmentation and classification networks for better wound care.