sennnnn / tissue-image-segmentation

Re-implementation of popular tissue image segmentation models

Tissue-Image-Segmentation

unofficial re-implementation of popular tissue image segmentation models

Support Model:

• UNet
• Dist
• DCAN
• MicroNet
• FullNet
• CDNet

Dataset Prepare

Please check this doc

Supported Dataset:

• MoNuSeg;
• CoNSeP;
• CPM17;
• CoNIC;

Installation

1. Install MMCV-full (Linux recommend): pip install MMCV-full==1.3.13;
2. Install requirements package: pip install -r requirements.txt;
3. Download tiseg: git clone https://github.com/sennnnn/Torch-Image-Segmentation;
4. Install tiseg: pip install -e .;

Usage

Training

# single gpu training
python tools/train.py [config_path]
# multiple gpu training
./tools/dist_train.sh [config_path] [num_gpu]
# demo (cdnet for CPM17 dataset on 1 gpu)
python tools/train.py configs/unet/unet_vgg16_radam-lr5e-4_bs16_256x256_7k_cpm17.py
# demo (unet for CPM17 dataset on 4 gpu)
./tools/dist_train.py configs/unet/unet_vgg16_radam-lr5e-4_bs16_256x256_7k_cpm17.py 4

Evaluation

# single gpu evaluation
python tools/test.py [config_path]
# multiple gpu evaluation
./tools/dist_test.py [config_path] [num_gpu]

Thanks

This repo follow the design mode of mmsegmentation & detectron2.

tissue-image-segmentation's Issues

why is the first channel of the sem_logit map set to 3?

When calculating loss, the truth value sem_gt is a matrix of size (256, 256) with values of 0 and 1, while sem_logit is a matrix of size (3, 256, 256). If there are only two classes, nuclei and background, why is the first channel of the prediction map set to 3?

link failure?

Git clone https://github.com/sennnnn/Torch-Image-Segmentation this link failure, I want to ask which '{:s}/weight_maps' MoNuSeg in CDN ent file come from? Is there any relevant processing code?