FAF GA Segmentation Pipeline

This repository contains a pipeline for classifying and segmenting geographic atrophy (GA) from FAF images.

The project includes:

• ga_classifier/: FAF image classification model and inference script
• ga_segmentation/: SAM-based segmentation inference using a pretrained model
• models/: pretrained model weights for classification and segmentation
• Dockerfile and docker-compose.yml: containerized environment for reproducible execution
• run_pipeline.sh: wrapper script that runs classification followed by segmentation

Docker usage

Build the container:

docker compose up --build

Repository structure

Important files to download separately:

• models: folder to create to store models
• models/classification_model_weights.ckpt: classification model weights
• models/segmentation_model_weights.pth: segmentation model weights
• imgs/: folder where images can be placed - if images are somewhere else, mount the location in the docker-compose file

Important input file:

• FAF_files.csv: example input CSV file (this is the default name, change it in the docker-compose)

Other files:

• ga_classifier/classify.py: main classification script
• ga_classifier/classify.sh: sample classifier launch script
• ga_classifier/tensors/: cache directory for preprocessed image tensors
• ga_segmentation/GASegPureInference.py: segmentation inference script
• ga_segmentation/segment_ga.sh: sample segmentation launcher
• run_pipeline.sh: example full pipeline invocation
• results/: expected output directory for segmentation masks

Notes

• The classification script caches preprocessed image tensors under ga_classifier/tensors/.
• Segmentation outputs are written to results/ga_segmentations/ by default.
• If your environment does not have a GPU, update the --gpu argument in run_pipeline.sh and the corresponding script variables to use CPU-only execution.
• Verify that the expected model weights exist in models/ before running inference.