Study of Chronic Wound Image Segmentation: Impact of Tissue Type and Color Data Augmentation

Abstract

Chronic wound segmentation is an essential task for evaluating wound and its recovery progress. A physician usually measures a wound area to choose proper treatment according to wound conditions. However, precise measurement needs accurate image-region segmentation. With the advent of deep learning for semantic image segmentation, accuracy of region segmentation is dramatically higher than traditional methods. Unfortunately, semantic segmentation in prior work did not produce satisfactory outputs in wound image segmentation, even with a large training dataset. This work, therefore, rethinks about the challenge and aims at not only improving segmentation accuracy, but also studying the impact of wound tissue types and color on accuracy. Since an end-to-end approach of semantic segmentation in prior work performed relatively poorly, the proposed method employs both image processing and deep learning techniques. The experiments indicated that slough was the most challenging tissue to be segmented. Also, properly increasing color variety of wound images significantly improved segmentation performance. The accuracy of the proposed method was 72%, 40%, and 53% in terms of intersection over union for granulation, necrosis, and slough wound tissue types, respectively. The proposed method outperformed a prior end-to-end approach, even though this method employed particularly simpler neural network models and much smaller number of training images.