hrlblab / layer_segmentation

kidney layer segmentation

layer_segmentation

This project consists of a collection of CNN-based and Transformer-based models for comparison experiments in layer segmentation programs.

• Incorporating cross-species homologous data for collaborative training can enhance the performance of models in segmenting the cortex and medulla in human kidney histopathology images.

1. To train CNN-based models, including UNet, PSPNet, and Deeplab-v3+:

CUDA_VISIBLE_DEVICES=0 python train.py

To validate and get IoU and Dice score, change the path of weight and run:

python predict_img.py
python get_metrics.py

2. To train Transformer-based models, including TransUNet and Swin-UNet:

CUDA_VISIBLE_DEVICES=0 python train.py --root_path '' --num_classes 5 --img_size 1024

To validate and get IoU and Dice score, change the path of weight and run:

python predict_img.py
python get_metrics.py

By utilizing external homologous data, the models have become better at perceiving edge textures, performing better in more precise localization of kidney layer boundaries.