Evaluation of Heat Map Methods Using Cell Morphology for Classifying Acute Lymphoblastic Leukemia Cells

Abstract

Explainable artificial intelligence (XAI) is a field of research that has attracted the interest of researchers in recent years. These algorithms seek to provide transparency to artificial intelligence (AI) models. One application of these algorithms is in the medical area, created as an auxiliary tool for corroborating predictions obtained by an AI when classifying pathologies, for example, Acute Lymphoblastic Leukemia (ALL). The present work evaluates the amount of information heat maps provide and how they relate to the blood components' morphological characteristics. For the assessment, four Convolutional Neural Network (CNN) models were retrained and fine-tuned to classify unsegmented images (ALL_IDB2 database). Subsequently, their respective heat maps were generated with the LRP (Layer-wise Relevance Propagation), Deep Taylor, Input*Gradient, and Grad-Cam methods. The best results were obtained with the GoogleNet model and the Grad-Cam heat map generation method, having a percentage of 43.61% of relevant pixels within at least one cell morphological feature present. Moreover, the most significant pixels are within the nucleus, with 73.97% of important pixels inside. According to the results, the Grad-Cam method best relates the relevant pixels generated in the heat map to the morphology of the cell of interest to classify a healthy or diseased cell.