Rabies diagnosis in low-data settings: A comparative study on the impact of data augmentation and transfer learning

A research team led by Khalil Akremi and Ines Abdeljaoued-Tej has developed an AI system that automates rabies diagnosis from fluorescent microscopy images, addressing critical shortages of skilled laboratory personnel in many African and Asian countries. The system employs transfer learning with four deep learning architectures—EfficientNetB0, EfficientNetB2, VGG16, and Vision Transformer (ViTB16)—and was trained on a remarkably small dataset of just 155 microscopic images (123 positive, 32 negative). To overcome data limitations, the team rigorously evaluated three data augmentation strategies, finding that TrivialAugmentWide best preserved critical fluorescent patterns while improving model robustness.

The EfficientNetB0 model, enhanced with Geometric & Color augmentation and validated through stratified 3-fold cross-validation, achieved the best classification performance on cropped images. Despite challenges like severe class imbalance and minimal training data, the research confirms the viability of deep learning for this high-stakes medical task. The team has already deployed an online tool to facilitate access, establishing a practical framework that could transform rabies diagnostics in low-resource settings and improve public health surveillance where traditional gold-standard methods are often unavailable.