Feature extraction procedures for chronic wound tissue classification

As healthcare services become both more necessary and expensive on a daily basis, the possibility arises for the inclusion of (semi)automated machine learning based approaches for aiding healthcare professionals and, hopefully, reducing costs. The area of chronic wound management, which includes diagnostics, healing process tracking, treatment related decision-making etc., represents a prominent candidate for such applications. This paper considers a straightforward approach to the classification of different tissue types constituting chronic wounds. It delves into the merits of several different feature extraction procedures and their descriptive properties when utilised at the pixel level, with the goal of selecting simple descriptors yielding solid performance and robust model behaviour. A multifaceted comparative analysis was conducted which considered performance on data stemming from both a prosthetic model and real patients in terms of different feature extraction procedures, colour spaces and window sizes. Even though the best advice anyone can ever offer is to adapt to the data at hand, the results do point at some possible shortcuts regarding feature extraction that lead to overall good performance figures.