FALCON: FoveA LoCalizatiON in En-face OCT Imaging via Explainable B-scan Classification and Transformer-Based Segmentation
This repository is the official implementation of FALCON: FoveA LoCalizatiON in En-face OCT Imaging via Explainable B-scan Classification and Transformer-Based Segmentation. And also a part of 42-687 Projects in Biomedical AI (Spring 2025) at Carnegie Mellon University.
Members: Arav Jain, Kitiyaporn Takham, Micah Baldonado, Shreyas Sanghvi
folder: fovea_classification_codes
The VGG16 was fine-tuning with fovea/non-fovea B-scan slices in both healthy and unhealthy eyes.
folder: fovea_explainability_codes
The fovea explainability algorithms include:
- Vanilla Seliency
- Occlusion
- GradCAM
- GradCAM++
- ScoreCAM
folder: enface_segmentation_codes
The model consists of ViT-Based MAE (encoder) and UNet++ Decoder (decoder).
folder: fovea_classification_codes
The model was evaluated on a sample volumetric
- potential fovea images (15 slices)
- potential non-fovea images (117 slices)
You can download pretrained models here:
- Fovea Classification trained on VGG16 using learning rate = 0.0001, Adam optimizer, and 10 epochs
Our fovea classification model achieves the following performance on fovea_classification_codes folder:
| Model name | Potential Fovea | Potential Non-fovea |
|---|---|---|
| best_VGG_model_1 | 80% | 96.58% |