MedSAM OCT Layer Segmentation — Inference

Fine-tuned MedSAM model for segmenting retinal layers (RNFL, GCL, IPL) in OCT B-scans.

---

Setup

pip install torch transformers Pillow numpy pandas matplotlib tqdm

---

Usage

1. Open run_inference.sh and set:

   • INPUT — path to an image, a directory of images, or a .csv file
   • OUTPUT — where to save results (default: ./output)

2. Run:

bash run_inference.sh

---

Input formats

Type	Example
Single image	INPUT="/data/scan.png"
Directory	INPUT="/data/oct_bscans/"
CSV file	INPUT="/data/scans.csv"

Supported image formats: .png, .jpg, .jpeg, .tif, .tiff, .bmp

If using a CSV, it must have a column containing image paths. The default column name is file_path — change --image_col in run_inference.sh if yours differs.

---

Outputs

For each image, the following are saved to OUTPUT:

File	Description
<stem>_mask.npy	Binary segmentation mask as float32 array
<stem>_vis.png	Overlay visualization
inference_results.csv	Per-image paths summary
inference_summary.json	Run metadata

---

Advanced options

Pass these directly to medsam_inference.py if needed:

Flag	Default	Description
--device	0	GPU index; use -1 for CPU
--threshold	0.5	Sigmoid threshold for binary mask
--image_col	file_path	CSV column containing image paths
--no_vis	off	Skip saving visualization PNGs