<![CDATA[Background: Leukemia is a malignant hematological disorder characterized by abnormal proliferation of white blood cells (WBCs), affecting over 474,000 patients globally each year, with manual diagnosis limited by inter-observer variability of up to 40%. Objective: This study evaluates whether U-Net-based WBC segmentation improves leukemia classification by comparing three configurations: VGG16 on original images, VGG16 with U-Net-predicted masks, and VGG16 with ground truth masks. Methods: A total of 3,256 ALL microscopic blood smear images were divided into training, validation, and testing sets (70:10:20). U-Net was trained using a combined Binary Cross-Entropy and Dice loss, and the segmentation outputs were used as inputs for VGG16 classification. Performance was evaluated using DSC, IoU, accuracy, precision, recall, F1-score, and AUC-ROC. Results: U-Net achieved a DSC of 0.7885 and an IoU of 0.7134. VGG16 with U-Net masks achieved 98.62% accuracy (AUC 0.9958), while VGG16 with ground truth masks achieved 98.31%. The baseline VGG16 achieved the highest performance with 99.69% accuracy (AUC 0.9999). Conclusion: Explicit segmentation did not improve classification performance for the dataset used in this study and may introduce unnecessary computational overhead when discriminative features are already preserved in the original images. Suggestion: Future studies should evaluate heterogeneous datasets and advanced segmentation architectures to further investigate the contribution of segmentation to leukemia classification.]]>
Copyrights © 2026
