Blood Cell Image Segmentation Using Convolutional Decision Trees and Differential Evolution

Semantic segmentation is an important process in computer vision that assigns labels to the pixels of an image to divide it into regions of interest. The most used machine learning model for this problem is the Convolutional Neural Network (CNN), in which high-performance results are obtained, however, they are difficult to understand and explain, which is not very useful in fields where explainability is fundamental, as in medicine. As an alternative, there are Convolutional Decision Trees (CDT), a tool that is easy to interpret due to its intuitive and user-friendly graphic structure. In this article, a method is proposed to induce an optimized CDT with different kernel sizes using the Differential Evolution algorithm, obtaining F1-scores greater than 0.92 on a set of blood cell images for erythrocyte segmentation, a relevant task for doctors and laboratory technicians.