A Robust and Explainable Deep Learning Pipeline with Hybrid Image Enhancement for Sickle Cell Detection

Abstract

In this study, a hybrid image enhancement and explainable deep learning framework is proposed for the automatic detection of morphological features associated with sickle cell disease in microscopic peripheral blood smear images. The experimental dataset comprises 991 images and exhibits a pronounced class imbalance. To address this limitation, data integrity was strengthened through systematic duplicate and near-duplicate removal, followed by the implementation of a leakage-safe data splitting strategy. Preprocessing strategies were comprehensively evaluated to mitigate variations in contrast, illumination, and staining conditions, with particular emphasis on a controlled contrast enhancement approach. In parallel, multiple methodological components, including augmentation ablation experiments, a three-phase dynamic learning strategy, validation-based threshold optimization, region-of-interest masking, and weighted ensemble techniques, were systematically examined. Model performance was assessed using standard evaluation metrics, including ROC-AUC, accuracy, sensitivity, specificity, and F1-score, and further analyzed through Grad-CAM-based explainability and quantitative attention analysis. The results indicate that controlled contrast enhancement significantly improves discriminative performance; however, the selection of an appropriate operating point remains dependent on clinical priorities. Overall, the proposed framework provides a more robust and clinically interpretable decision-support pipeline for binary classification of microscopic images, highlighting the importance of data leakage prevention, error distribution analysis, and explainability in biomedical image analysis.