Extend RETFound to OCT Drusen Segmentation with Lightweight Decoder

[MDSALMANSHAMS]

This PR extends the RETFound foundation model to support pixel-level segmentation tasks.

Specifically, we demonstrate RETFound-based drusen segmentation on the MendeleyOCT dataset by attaching a lightweight convolutional decoder to the pretrained ViT encoder.

Key contributions:

• Added a lightweight segmentation decoder (models_segmentation.py) that converts ViT patch embeddings to full-resolution masks.
• Implemented a training pipeline for segmentation (main_segmentation.py) using a combined CrossEntropy + Dice loss.
• Added reusable training utilities (engine_segmentation.py) for segmentation tasks.
• Provided a standalone inference script (inference_segmentation.py) for evaluation and overlay visualization.
• Included an example notebook (examples/RETFound_MendeleyOCT_demo.ipynb) demonstrating dataset preparation, training, and inference.

Dataset notes:

• OCT images were horizontally cropped (top and bottom) to focus on the retinal region.
• Binary segmentation masks represent drusen regions.

The implementation keeps the original RETFound encoder unchanged and adds only a minimal decoder, maintaining compatibility with pretrained weights.

This demonstrates the adaptability of RETFound representations for dense prediction tasks beyond classification.