INVESTIGATING MYELOID AND LYMPHOID BLAST CLASSIFICATION USING PERIPHERAL BLOOD SMEARS

Abstract

To diagnose a patient with Acute Myeloid Leukemia or Acute Lymphoid Leukemia, experts manually inspect peripheral blood or Bone Marrow Aspirate smears to obtain myeloblast and lymphoblast counts. These expert analyses are prone to errors, in some cases leading to 64.8-70.5% intra-observer and 63-72% inter-observer concordance between experts, and thus immunophenotyping and a cytochemistry study are typically performed concurrently to guarantee an accurate estimate of blast counts. In this paper, we develop a computer vision model in pursuit of understanding whether these blast counts can be derived con-sistently in peripheral blood smears, which could prevent the need for additional analyses. Our model extends a pre-trained ResNeXt-50 backbone, resulting in an AUC greater than 0.969 for all classes, with F1-scores of 0.737, 0.889, and 0.941 for the Lymphoid, Myeloid, and Other classes, respectively, which matches the performance of experts. We establish that these results can serve as a baseline for future research, and, using the explainable AI method XGrad-CAM, simultaneously show that the concordance between experts might be due to visual overlap of myeloblasts and lymphoblasts. Finally, we discern the weaknesses of the model, and propose mitigation strategies to reduce their ramifications.

Date of Award	19 Aug 2025
Original language	English
Supervisor	Lyana Curier (Examiner) & Daniel Tan (Co-assessor)