Explainable Early Detection of Diabetic Foot Ulcers Using Thermal Imaging with Vision Transformers and Grad-CAM

This project utilizes Vision Transformer (ViT) and Data-efficient Image Transformer (DeiT) models to classify plantar thermogram images for the early detection of Diabetic Foot Ulcers (DFU).

To validate model performance, I employed Stratified 5-Fold Cross-Validation for robust evaluation across balanced data splits and utilized a Class-Weighted Loss function to address class imbalance. Furthermore, Grad-CAM visualizations are integrated to provide explainable AI insights into the model's diagnostic regions.

Huggingface space | Dataset

GradCam Visulization

I use Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the regions of the thermogram that the models focus on when making a prediction.

ViT Attention Map	DeiT Attention Map

Model Performance

The following table shows the performance metrics (Mean ± Std) across folds for the Vision Transformer (ViT) and Data-efficient Image Transformer (DeiT) models.

Model	Accuracy	Precision	Recall	F1 Score
ViT	0.9193 ± 0.0489	0.9416 ± 0.0493	0.9508 ± 0.0471	0.9452 ± 0.0330
DeiT	0.9221 ± 0.0387	0.9541 ± 0.0180	0.9385 ± 0.0456	0.9459 ± 0.0277

Visualizations

Confusion Matrix

Comparison of classification performance between ViT and DeiT.

Vision Transformer (ViT)	Data-efficient Image Transformer (DeiT)

Training History

Average Accuracy and F1 Score across folds over training epochs.

Vision Transformer (ViT)	Data-efficient Image Transformer (DeiT)