Install PyTorch 1.9 then pip install . at the root of this repository.
We release models with a Vision Transformer backbone initialized from the improved ViT models.
Double Layer Segmentation models with ViT backbone:
| Download | |
|---|---|
| model | config |
You can generate segmentation maps from your own data with:
python -m segm.inference --model-path models/checkpoint.pth -i data/images/ -o segmaps/ The color coding rule for the segmentation mask is defined as DLS[0,255,255], Drusen[255,255,0], Background[128,0,255].
During the inference, all the images are rescaled to 512x512 then obtained prediction masks with the same resolution will be rescaled back to 500 x 768.
500(w) x 768(h), rgb(24bit per pixel, but grayscale) with JPG format
The resolution of the original B-scan image was 500 pixels in width, and 1536 pixels in height. The B-scan was then systematically cropped into 500x768 accordingly so that the retinal layer comes to the center of the cropped image.
500(w) x 768(h), rgb(24bit per pixel) with JPG format
Usage:
python crop_bscan.py --inputdir [path to the original B-scan (500x1536) directory] --outputdir [path to the output directory]
Usage:
python enface_proj.py --inputdir [path to the directory where the prediction masks are stored] --outputdir [path to the directory you want to store projection maps] --eyeid [specify eyeID e.g. "P2202_OS"]
@article{kihara2022detection,
title={Detection of Nonexudative Macular Neovascularization on Structural OCT Images Using Vision Transformers},
author={Kihara, Yuka and Shen, Mengxi and Shi, Yingying and Jiang, Xiaoshuang and Wang, Liang and Laiginhas, Rita and Lyu, Cancan and Yang, Jin and Liu, Jeremy and Morin, Rosalyn and others},
journal={Ophthalmology Science},
volume={2},
number={4},
pages={100197},
year={2022},
publisher={Elsevier}
}
The Vision Transformer code is based on timm library and the semantic segmentation pipeline is using mmsegmentation and segmenter.