GA Classifier Inference

Inference scripts for the three binary geographic atrophy (GA) classifiers: FAF, SLO, and OCT.

Scripts

File	Description
infer_2d.py	Inference script for 2D modalities (FAF and SLO)
infer_oct.py	Inference script for OCT volumes (MIL-based)
infer_2D.sh	Config + launcher for FAF and SLO inference
infer_oct.sh	Config + launcher for OCT inference

Usage

bash infer_2D.sh
bash infer_oct.sh

To change any parameter (checkpoint path, image directory, threshold, etc.), edit the corresponding .sh file directly.

Arguments

infer_2d.py (FAF / SLO)

Argument	Default	Description
--ckpt_path	required	Path to Lightning checkpoint (.ckpt)
--image_dir	required	Directory of input .png images
--output_csv	required	Path to write predictions CSV
--threshold	0.5	Classification threshold on sigmoid probability
--image_size	512	Image resize dimension (square)
--batch_size	32	Images per batch
--gpu	0	CUDA device index (CUDA_VISIBLE_DEVICES)

infer_oct.py (OCT)

Argument	Default	Description
--ckpt_path	required	Path to Lightning checkpoint (.ckpt)
--volume_dir	required	Directory of volume subfolders (each containing .png slices)
--output_csv	required	Path to write predictions CSV
--threshold	0.5	Classification threshold on sigmoid probability
--image_size	256	Slice resize dimension (square)
--num_frames	49	Number of B-scan slices per volume (pads or truncates to this)
--batch_size	12	Volumes per batch
--gpu	1	CUDA device index (CUDA_VISIBLE_DEVICES)

Input Format

• FAF / SLO: flat directory of .png files
• OCT: directory of volume subfolders, each containing .png B-scan slices named in sorted order (e.g., slice_000.png, slice_001.png, ...)

Output Format

All scripts write a CSV with the following columns:

Column	Description
filepath	Absolute path to the input image or volume folder
volume_name	(OCT only) Name of the volume subfolder
probability	Sigmoid output of the model (0–1)
prediction	Binary label: 1 = GA+, 0 = GA−

Models

Modality	Architecture
FAF	ResNet50 + dropout head
SLO	ResNet50 + dropout head
OCT	ResNet18 feature extractor + gated attention MIL