Multi-Scale Attention Network for Diabetic Foot Ulcer Segmentation using Self-Supervised Learning

Overview

This project focuses on detecting diabetic foot ulcers using a self-supervised learning approach combined with attention mechanisms. The model utilizes a pretrained Vision Transformer (ViT) from DinoV2 and multi-scale DenseNet for feature extraction, followed by an Attention UNet for segmentation.

Scripts

• main.py: Trains the self-supervised model using unlabelled images to extract meaningful features.
• fine_tune.py: Fine-tunes the pre-trained model for segmentation tasks using labeled data.
• inference.py: Runs inference on new images to generate ulcer segmentation masks.

Results

• Self-Supervised Learning: MSE Loss = 0.01
• Segmentation (Attention UNet):
  • BCEWithLogits Loss = 0.14
  • Dice Coefficient Loss = 0.3054

Segmentation Results

Instructions

1. Replace Paths: Update the paths in the scripts with your own paths.
2. Run the Scripts: Execute the scripts (main.py, fine_tune.py, inference.py) in sequence for training, fine-tuning, and inference.

Authors

• Aravind shrenivas Murali - Graduate Student, University of Arizona, [email protected]
• Marino Alejandro Chuquilin Fernandez - Undergraduate Student, University of Arizona, [email protected]
• Dr. Eung-Joo Lee - Assistant Professor, University of Arizona, [email protected]

License

This project is licensed under the MIT License.