An integrated image segmentation framework for white blood cell detection in microscopic blood smears: A hybrid approach for hematology diagnostics

Abstract

White blood cells (leukocytes) are fundamental components of the immune system, and their enumeration and morphological assessment via microscopic image analysis serve as critical diagnostic indicators. However, automatic segmentation of white blood cells remains challenging due to variations in staining, illumination, and complex cellular morphology. This study proposes a novel hybrid segmentation framework that integrates GrabCut and Otsu thresholding methods. The proposed framework addressing two fundamental limitations of previous approaches: (1) GrabCut incorporates inter-pixel spatial relationships through graph-based optimization, which traditional thresholding methods ignore, while (2) Otsu thresholding provides automatic, data-driven binarization that eliminates manual parameter tuning required by standalone GrabCut. Their sequential integration, where GrabCut refines foreground extraction followed by Otsu for adaptive binarization, enables the framework to simultaneously optimize color discrimination, edge preservation, and segmentation consistency under varying image conditions. This synergy produces more robust segmentation than either method applied independently. The proposed framework was evaluated on three datasets: primary data (author-collected), JTSC, and CellaVision. Results demonstrate dataset-dependent performance trade-offs. On primary data, the framework achieved 94.55% accuracy, though precision, recall, F1-score, and IoU remained low. On JTSC data, it achieved efficient computation (0.06 seconds per image), good accuracy (92.07%), and moderately high precision (0.8467) but low recall (0.4860). On CellaVision data, it achieved high accuracy (93.97%) and excellent precision (0.9495) but low recall (0.4796). These findings indicate that the proposed framework successfully improves precision and computational efficiency while maintaining high accuracy; however, recall limitations persist, suggesting that the current combination is better suited for applications prioritizing detection reliability over complete sensitivity.