Data Processing

Inference and Evaluate Using Segmentation Model

Please refer to MIC-DKFZ/nnUNet to build the segmentation environment and run the code.

The last iteration(test on Messidor) of the segmentation model(nnU-Net) in Fundus-Engine is open sourced to huggingface for reference. You can refer to infer_seg.py to call the nnU-Net model for inference.

You can refer to eval_seg.py to evaluate. Usage:

python eval_seg.py \
    --folder_ref gt/folder \
    --folder_pred predicted/folder \
    --output_dir output/dir

Convert Segmented Pixels to Bbox Coordinates

pixel2bbox.py processes binary segmentation masks to generate bounding boxes for objects of interest.

Usage:

python pixel2bbox.py \
    --mask_folder /path/to/masks \
    --image_folder /path/to/images \
    --output_folder /path/to/output \
    --label_name "ObjectOfInterest" \
    --pixel_intensity 1 \
    --eps 160 \
    --min_samples 10 \
    --area_threshold 100

The mask corresponding to the image should have the same file name.

Label_name is the name of the feature corresponding to the mask.

Pixel_intensity, eps, min_samples and area_threshold are clustering parameters.