Skip to content

Latest commit

 

History

History
29 lines (15 loc) · 1.77 KB

README.md

File metadata and controls

29 lines (15 loc) · 1.77 KB

Lesion Detection Network

A wonderful object detection toolbox MMDetection are used for lesion detection. A trained model and predicted results can be downloaded here. Note that the model has a relatively poor generalization ability and cannot precisely predict lesions of fundus images from EyePACS because of the limited training samples of IDRiD.

Usage

Installation

Please refer to the instruction of MMDetection for installation.

Training

  1. Download IDRiD dataset [link].

  2. Use scripts/mask2bbox.py to convert the lesion mask of images to bounding boxes. The output is a csv file.

  3. Use scripts/csv2coco.py to convert the csv file to coco annotation format. The requirements for the format of new datasets can be found in here.

  4. Replace the configuration files in MMDetection with the files in configs. Please remember to update the path to dataset.

  5. Follow the instruction of MMDetection to training the model.

Dataset for Representation Learning

  1. Download EyePACS dataset [link]. Then delete line 77 in ../tools/crop.py and run it to remove the black border of images of EyePACS. Line 77 in ../tools/crop.py is for resizing and we do not resize images for lesion detection.

  2. Use scripts/detect.py to generate lesion predictions of EyePACS.

  3. Use scripts/build_dataset.py to select lesions with high confidence threshold. The output is a pickle file for training the lesion-based contrastive learning models.