Automated blood cell quantification for disease forecasting

Abstract

Peripheral blood smear analysis is a critical diag-nostic tool for identifying infections, anaemia, and hematologic malignancies. However, manual microscopy is time-consuming and susceptible to inter-observer variability, rendering it imprac-tical for large-scale clinical applications. Recent advancements in deep learning models, such as DeepLabV3+, have demon-strated potential in automating medical image segmentation. Nevertheless, many existing models lack robustness in handling noisy images and fail to integrate diagnostic predictions based on cell counts. This study aims to develop a reliable, end-to-end system for the segmentation, quantification, and inference of blood cell-related diseases. We employ DeepLabV3+ with a customised preprocessing pipeline to accurately segment white blood cells (WBCS), red blood cells (RBCS), and platelets, while inferring conditions such as leukaemia or infection based on established count thresholds. Our model attained a high segmentation accuracy, achieving an Intersection over Union (IoU) of up to 0.98, facilitating automated and interpretable clinical insights. This research supports scalable, AI -assisted haematological screening and contributes to reducing diagnostic workload within under-resourced healthcare systems.