Regionally focused neural-coder model designed for the diagnosis of acute lymphoblastic leukemia disease

Abstract

Cancer ranks as the second leading cause of death worldwide, accounting for one in six deaths. Early diagnosis of cancer in patients significantly increases their survival chances by accelerating the treatment process. One of the most prevalent types of cancer today is leukemia, also known as blood cancer. The diagnosis of leukemia involves examining bone marrow smears and peripheral blood smears under a microscope in a laboratory setting, with detailed observation of these two indicators. However, specialists in this field may not always be available, leading to potential human errors. In this study, 3256 publicly available PBS images were used to diagnose Acute Lymphoblastic Leukemia (ALL). Initially, the images were trained using the DarkNet19 ESA model, and deep features of the images were extracted. Then, features were obtained by training the images with the GradCAM algorithm and combining these features with those from the DarkNet19 ESA model. The features were further trained using AutoEncoder networks, doubling the number of features. Finally, effective features were identified using the Harris Hawk optimization algorithms and classified with machine learning algorithms to diagnose ALL from PBS images. As a result, the disease was diagnosed with an accuracy rate of 98.52%. The proposed model enables the early and rapid detection of ALL.