Deep Learning for Automating the Immunophenotyping Assessment in Childhood Acute Leukemia Diagnosis

Abstract

Acute leukemias are a heterogeneous group of hematologic malignancies characterized by rapid progression, making early and accurate diagnosis critical for effective treatment. These diseases are primarily classified based on the lineage of the affected precursor cells, which begin to proliferate uncontrollably. Immunophenotyping using Multiparameter Flow Cytometry (FCM) is a key diagnostic tool [1], which provides high-dimensional, multi-parameter analysis of individual cells essential for distinguishing between the lineages and phenotypes of childhood acute leukemia. However, immunophenotyping evaluation remains a labor-intensive and expert-driven process, making it susceptible to subjectivity and variability across clinical settings. To address these challenges, we explore deep learning methodologies based on self-attention mechanisms to automate and enhance immunophenotyping assessment. Our approach is based on transformer-derived architectures tailored to unordered, high-dimensional tabular data, making them particularly well-suited for modeling flow cytometric data where feature ordering holds no intrinsic significance. In particular, we refine the Set Transformer framework [2] to better capture complex feature dependencies while preserving permutation invariance, aligning naturally with the structure of flow cytometric data. Additionally, we integrate specialized activation functions that enhance representational capacity and optimization stability, resulting in improved classification performance compared to previous deep learning approaches [3]. The model is trained in a supervised manner, directly utilizing raw FCM data without requiring manual preprocessing, allowing it to fully leverage the richness of the high-dimensional input space. Experimental evaluations on a diverse dataset demonstrate that our approach effectively distinguishes between major lineages of pediatric acute leukemia, offering a promising step toward automating hematologic diagnostics. The results indicate that leveraging self-attention enhances both classification accuracy and generalization across heterogeneous patient samples. This study underscores the potential of deep learning-driven immunophenotyping to complement traditional diagnostic workflows, paving the way for more efficient and reproducible leukemia assessments in clinical practice.