Feasibility study of an AI-powered mobile app to support cutaneous leishmaniasis diagnosis in the Brazilian Amazon

Cutaneous leishmaniasis (CL) remains a major public health challenge, especially in Brazil's Amazon, where environmental and economic pressures sustain transmission. Delayed diagnosis drives morbidity, stigma, and costs. Community health workers are pivotal yet under-equipped for early triage. Artificial intelligence, effective in dermatologic imaging, is underused for CL; feasible, offline clinical tools could accelerate referral and timely care. This study aimed to evaluate the feasibility of an AI-assisted tool for early CL triage in the Brazilian Amazon by: (i) developing an AI model for CL identification from clinical skin lesion images; (ii) integrating the model into an offline-optimized mobile application for resource-limited settings; and (iii) conducting initial, real-world clinical validation.

Exploratory, IRB-approved feasibility study. Retrospective images from Brazilian Amazon sites informed a two-stage AI pipeline (lesion segmentation+classification) integrated into an offline mobile app. Multicenter validation in ideal- and real-world scenarios. Primary metric AUC-ROC; secondary sensitivity/specificity. One-sided tests: AUC-ROC [Formula: see text]0.70; ideal-world sensitivity [Formula: see text]0.75.

For the Classification Model, 64 images were assigned to the test set while 1,160 were used for training and validation (80:20 split), with DenseNet121 yielding the highest accuracy of 0.88. The full AI pipeline (Segmentation and Classification models) achieved an accuracy of 0.81, an F1-score of 0.80, an AUC-ROC of 0.90, and a sensitivity of 0.76. In the ideal-world analysis, sensitivity reached 0.92, the F1-score was 0.84, and specificity was 0.42.

We demonstrate the feasibility of an offline, AI-assisted mobile tool to support triage and referral for cutaneous leishmaniasis in the Brazilian Amazon. Performance reflects preliminary, descriptive point estimates from an initial diagnostic accuracy assessment and should be interpreted with caution; the tool is not intended for standalone diagnosis. Next steps include prospectively powered clinical validation, usability refinements, and regulatory evaluation, including alignment with ANVISA requirements in Brazil. Overall, this work represents an initial step toward closing the gap in clinically supported diagnostic tools for neglected tropical diseases in resource-constrained settings.