Deep Feature Selection and AutoML Ensembles for Acute Lymphoblastic Leukemia Cell Classification

Abstract

Acute lymphoblastic leukemia (ALL) classification from microscopic peripheral blood smear images remains challenging because of staining variability, inter-class morphological similarity, and class imbalance. This paper presents a four-class classification framework for distinguishing benign hematogones, Early Pre-B ALL, Pre-B ALL, and Pro-B ALL, rather than the easier binary benign-versus-malignant screening task. The framework uses VGG19 as a fixed deep feature extractor, Mutual Information (MI) for label-aware feature selection, and H2O AutoML stacked ensembles for automated classifier construction. On the held-out 80:20 test partition, the MI-based configuration achieved 94.14% overall accuracy with a Wilson 95% confidence interval of 92.02%-95.72% and a macro-F1 of approximately 0.931. Class-level analysis showed strong performance for Pre-B and Pro-B cells, while benign hematogones remained the most difficult class because of visual similarity with early malignant cells. The results support the value of combining deep feature embeddings with automated model selection for ALL image classification, but the system is not yet a substitute for clinical diagnosis without patient-level and multi-center external validation.