A Lightweight Deep Learning Framework for Reliable Microscopy-Based Diagnosis of Cutaneous Leishmaniasis

Purpose Cutaneous leishmaniasis (CL) is a neglected tropical and zoonotic disease that threatens both human and animal health. Microscopic diagnosis remains the reference standard but is time-consuming and operator-dependent. This study aimed to develop a lightweight, calibration-aware deep learning framework for automated amastigote detection and slide-level diagnosis from Giemsa-stained microscopy images. Methods A U-Net architecture with a MobileNetV2 encoder was implemented for pixel-level parasite segmentation. A weakly supervised pseudo-labeling strategy was used to generate training masks from limited annotations. Post-hoc probability calibration was applied using isotonic regression, Platt scaling, and temperature scaling. Model performance was assessed through Dice coefficient, Intersection-over-Union (IoU), AUROC, Brier score, and Expected Calibration Error (ECE) metrics on an independent test set. Results The proposed model achieved a Dice coefficient of 0.905 and IoU of 0.826 for segmentation, with an AUROC of 0.986 for diagnostic discrimination. Isotonic calibration improved predictive reliability, reducing the Brier score to 0.0240 and ECE to 0.0229 without affecting AUROC or AUPRC values. Statistical validation confirmed consistent calibration and discrimination across all methods. Conclusion Isotonic calibration enhances the interpretability and reliability of deep-learning-based CL diagnostics. The proposed calibrated framework provides a robust and cost-efficient diagnostic solution for zoonotic diseases, supporting early detection and control within the One Health context. Keywords: Cutaneous leishmaniasis; Deep learning; U-Net; MobileNetV2; Isotonic regression; Microscopy; One Health