Abstract:
Early and accurate diagnosis of monkeypox is critical, especially in resource-limited
settings where access to laboratory diagnostics like PCR is constrained. This study explores
the integration of transfer learning and optimization strategies using lightweight
convolutional neural networks (CNNs), specifically MobileNetV2 and EfficientNetB0,
for the classification of monkeypox, chickenpox, measles, and normal skin lesions.
Multiple training configurations were implemented using two optimizers (Adam and
SGD), two learning rates (0.001 and 0.0001), and four class imbalance handling strategies
(none, class weights, oversampling, both). Results show that MobileNetV2 consistently
outperformed EfficientNetB0, with feature extraction and class weights under
Adam optimizer at a 0.001 learning rate achieving the highest accuracy (85.00%)
and AUPRC (0.9284). Grad-CAM was integrated to enhance interpretability, offering
real-time visual explanations of model predictions. The best-performing model
was deployed in a React Native mobile application with a Flask backend, capable
of real-time image classification and explainability. This study demonstrates the
feasibility and clinical relevance of deploying interpretable, lightweight CNN models
for mobile-based monkeypox diagnosis. The final application, GabAI: Grad-Aided
Bioscan Intelligence, showcases how Explainable AI can be deployed in mobile platforms
to support clinical decision-making.
