Reliable Leukemia diagnosis and localization through Explainable Deep Learning
Marcello Di Giammarco, Benedetta Dukic, Fabio Martinelli, Mario Cesarelli, Fabrizio Ravelli, Antonella Santone, Francesco Mercaldo · 2024
Acute lymphoblastic leukemia is a cancer of the blood and bone marrow, it is the most common form of childhood cancer. In the United States, approximately 75% of people under age 20 diagnosed with leukemia are diagnosed with acute lymphoblastic leukemia. An estimated 400 people ages 15 to 19 in the US are diagnosed with the disease each year. In this paper we propose a Deep Learning network-based approach to detect acute lymphoblastic leukemia from images of blood cells. Additionally, the suggested approach can offer predictability through class activation mapping, which aims to automatically highlight the relevant and suspicious patterns in the image. We consider a method that uses the output of two separate class activation mapping techniques to determine whether the acute lymphoblastic leukemia prediction and localization can be regarded as resilient. Using a dataset of 6.099 blood cell images, we assess the efficacy of the suggested method and achieve an accuracy of $\mathbf{9 4 \%}$, demonstrating the usefulness of the proposed network for acute lymphoblastic leukemia detection and localization. Our method introduces also a similarity index aimed to “quantify” qualitative results coming from the heatmaps, in such a way as to improve the trustworthiness and reliability of the Artificial Intelligence for the medical staff.