High-Accuracy and Lightweight Image Classification Network for Optimizing Lymphoblastic Leukemia Diagnosisy.

Abstract

Leukemia is a hematological malignancy that significantly impacts the human immune system. Early detection helps to effectively manage and treat cancer. Although deep learning techniques hold promise for early detection of blood disorders, their effectiveness is often limited by the physical constraints of available datasets and deployed devices. For this investigation, we collect an excellent-quality dataset of 17,826 morphological bone marrow cell images from 85 patients with lymphoproliferative neoplasms. We employ a progressive shrinking approach, which integrates a comprehensive pruning technique across multiple dimensions, including width, depth, resolution, and kernel size, to train our lightweight model. The proposed model achieves rapid identification of acute lymphoblastic leukemia, chronic lymphocytic leukemia, and other bone marrow cell types with an accuracy of 92.51% and a throughput of 111 slides per second, while comprising only 6.4 million parameters. This model significantly contributes to leukemia diagnosis, particularly in the rapid and accurate identification of lymphatic system diseases, and provides potential opportunities to enhance the efficiency and accuracy of medical experts in the diagnosis and treatment of lymphocytic leukemia.