Deep Learning-Based Cross-Lineage Classification of Acute Leukemia Subtypes

The accurate sub classification of acute leukemia is critical for patient prognosis and treatment planning. However, it remains a challenging task. Manual diagnosis is labor-intensive and prone to human errors. Existing Artificial Intelligence (AI) enhanced automated methods typically focus on either broad leukemia categories or fine-grained subtypes within a single cell lineage (myeloid or lymphoid). This study addresses a significant gap by formulating a novel cross-lineage classification task and investigating the feasibility of a single, unified deep learning (DL) model to solve it. Unlike prior work, focusing on subtypes within a single lineage, we combine key subtypes from both Acute Myeloid Leukemia (AML) and Acute Lymphocytic Leukemia (ALL) thereby creating a challenging seven-class problem. A comparative analysis is performed on four prominent architectures: ResNet-50, DenseNet-121, InceptionV3, and VGG16, all trained from scratch on the public Raabin dataset. The experimental results show that InceptionV3 attained a superior performance by achieving an average accuracy of 99.51%±0.04%. Further class-wise analysis confirmed its robustness, with F1-scores exceeding 99% for all individual subtypes. These findings confirm that a single, unified model can effectively and accurately distinguish between fine-grained leukemia subtypes across different lineages. It leads towards a more comprehensive automated diagnostic tool.

Department

Electrical and Computer Engineering

Publisher

IEEE

Sponsor

Effat University; Universiti Tunku Abdul Rahman, Malaysia

Open Access

https://repository.effatuniversity.edu.sa/handle/20.500.14131/2808

DOI

10.1109/IICT68374.2026.11468462