Andrea Kaye A. Chiang, Lara R. Jakosalem, Christine D. Bandalan
Examining bone marrow smears enables the detection of hematologic diseases such as Acute Lymphoblastic Leukemia (ALL). However, manual diagnosis is time-consuming and susceptible to human error due to fatigue, varying expertise, and limited diagnostic tools. This study applies a combination of traditional image processing techniques (e.g., color space conversion, noise filtering, watershed segmentation) and deep learning models to improve ALL cell detection in raw bone marrow smear images collected from a local hospital. Raw bone marrow smear images, collected from a local hospital in Cebu, were enhanced using traditional methods and segmented using Cellpose, followed by classification using a CNN stacking ensemble. YOLOv5 was employed for cell detection and extraction. The pipeline achieved promising results, with an average segmentation IoU of 82.1%, Dice coefficient of 88.9%, and classification accuracy of 96.54%. The use of an unprocessed, locally sourced dataset highlights the practical relevance of this work compared to existing approaches. © 2025 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
