Diabetic foot ulcer risk detection

Diabetic foot ulcers are among the most expensive and most preventable complications of diabetes, and risk is routinely under-detected between clinic visits. A model that can flag risk from a smartphone image moves screening from the clinic to wherever the patient is.

A PyTorch pipeline using ResNet50 transfer learning for risk classification, with Grad-CAM explainability so clinicians can see which regions of the image drive each prediction rather than being asked to trust a score.

Cross-dataset generalisation testing checks that performance holds beyond the data the model was trained on. That is the failure mode that quietly kills most medical imaging models when they meet real-world images.

The pipeline is data-agnostic by design, so NHS data can be integrated under local governance without re-architecting anything. The work is conducted as an MSc dissertation (University of Reading) in clinical collaboration with Royal Berkshire Hospital.