Applying convolutional neural networks in identifying and classifying Acute Lymphocytic Leukemia and its subtypes using Matlab
Phan Nguyen Bao Huy, Tran Thien Hau, Tran Anh Tu · Journal of Physics Conference Series · 2025
Abstract Leukemia is a blood cancer originating in the bone marrow that leads to the production of abnormal blood cells, impacting cell functionality and efficiency. According to Globocan, statistics for 2022 show that Viet Nam has 180,480 new cancer cases and 120,184 deaths, of which leukemia has 5,789 new cases, ranking 8th among cancer types (accounting for 3.2%) and 4,330 deaths, ranking 6th in mortality rate among cancer types (accounting for 3.6%). Depending on the progression rate, leukemia can be identified as either acute or chronic. Acute leukemia affects blood cells in the early stages of development, causing the disease to become more severe, while chronic leukemia affects more developed blood cells, and the disease tends to develop slowly. Another way to classify leukemia is by the specific blood cell type it targets, either myeloid or lymphoid cells. Understanding the affected cell type is essential for diagnosis and treatment planning, as it influences disease behavior and response to treatment. This paper focuses on identifying and classifying ALL (Acute Lymphoblastic Leukemia) blasts in the most prevalent childhood cancer type using transfer learning of the Alexnet network model in Matlab software. Diagnosing ALL disease using peripheral blood smear (PBS) images helps screen and treat the disease early, contributing to improving treatment effectiveness for patients. Laboratory users often face challenges when examining PBS images, as the non-specific signs and symptoms of ALL can result in frequent misdiagnoses and diagnostic errors. The research presents a model for analyzing peripheral blood smear images, aimed at identifying ALL cell types and distinguishing among its malignant subtypes. The ALexnet model will be trained in 3 different ways including: Train with no parameters (Training a Neural Network from Scratch), train the pretrained network but change the fullyConnected layer at the end and train the pretrained network but change the 14th layer onwards. The results show that the model using the second training method achieves the best performance when reaching 98.664% accuracy on the validation set, while methods 1 and 3 are 92.292% and 89.825% respectively. Finally, the second model is saved to design the user application using AppDesigner.