An Artificial Intelligence Framework for Detection of Leukemia with Comparison Between Segmented and Non-Segmented Images.

Abstract

Background: Accurate classification of Acute Lymphoblastic Leukemia (ALL) subtypes from peripheral blood smear (PBS) images using deep learning may support automated image-based analysis of this blood malignancy.

Methods: We developed a custom lightweight Convolutional Neural Network (CNN) to classify white blood cells into four subtypes: Benign, Early Pre-B ALL, Pre-B ALL, and Pro-B ALL. We assessed performance using accuracy, ROC curves, and explainability techniques, including upsampled feature map activations, occlusion sensitivity, and SHAP-style explanation overlays, comparing original (Non-segmented) and HSV-Segmented images. To mitigate patient-level data leakage without explicit patient identifiers, we further used unsupervised K-means clustering of HSV color histograms to group 3256 images into 89 "proxy-patient" clusters. Consequently, we implemented rigorous subject-level grouped 5-fold cross-validation and Hold-Out protocols alongside standard image-level evaluations.

Results: Under standard image-level evaluation, the model achieved 98.8% accuracy on non-segmented images and 91.3% on segmented images. Crucially, this performance advantage was maintained under strict subject-level cross-validation (97.9% vs. 89.7%), confirming the robustness of the findings against data leakage. Explainability visualizations demonstrated that HSV segmentation discarded vital contextual morphological cues, increasing confusion between closely related subtypes.

Conclusions: A lightweight CNN can accurately classify ALL subtypes directly from raw PBS images. The evaluated HSV segmentation degraded performance by removing contextual information. This study underscores the necessity of subject-level evaluation in medical imaging and indicates that mandatory preprocessing should be empirically justified rather than assumed beneficial.