Fine-Grained Skin Wound Segmentation Based on Machine Learning with Scribble Annotations

In recent years, the increasing prevalence of chronic diseases and the aging population have intensified the demand for wound management. The assessment of different issue type in the wound area is critical for evaluating the healing process and making treatment plans. However, due to the complex compositions of the wound, it is difficult to obtain pixel-wise annotations that are required by machine learning methods for automatic fine-grained wound segmentation. In this context, this paper proposes a weakly supervised learning method based on scribble labels, which significantly reduces the labeling cost of wound segmentation. Based on limited training samples, multiple training sets are constructed using random sampling techniques, and multiple classifiers are employed for ensemble learning to improve the model's accuracy. Additionally, a post-processing technique based on superpixel algorithms is introduced to achieve smoother and more accurate segmentation results. Experimental results demonstrate that this method achieves reliable skin wound segmentation results, providing an innovative solution for skin wound analysis.