Qixiang ZHANG, Yi LI, Cheng XUE, Xiaomeng LI*, "Morphology-inspired Unsupervised Gland Segmentation via Selective Semantic Grouping", MICCAI 2023 (Accepted).
This MSSG framework is designed for unsupervised gland segmentation on histology images, containing two modules. The first Selective Proposal Mining (SPM) module generates proposals for different gland sub-regions. And the Morphology-aware Semantic Grouping (MSG) module groups the semantics of the sub-region proposal to obtain comprehensive knowledge about glands.
This code has been tested with Python 3.9, PyTorch 1.12.0, CUDA 11.3 mmseg 0.8.0 and mmcv 1.4.0 on Ubuntu 20.04.
Download GlaS dataset from Official Website, and place the dataset as following:
/your/directory/SPM/
└── glas/
├── training_images/
│ ├── xxxxxxxxx.bmp
│ └── ...
└── training_gts/
├── xxxxxxxxx.bmp
└── ...
Install Python library dependencies
pip install -r requirements.txt
Install MMSegmentation codebase, see documentation from MMLab for details.
You can simply download the pre-generated proposal map from Link to Proposal Map with extracted code rzgp
Or use the SPM module to generate candidate proposals for each histology image, and use the empirical cue to select gland sub-region proposals
cd SPM
python SPM.py --train_image_path ../glas/training_images --output_root result --nConv 3 --cluster_num 5
Crop the training image and the proposal map
cd MSG
python tools/crop_img_and_gt.py MSG/glas/images SPM/proposal_map MSG/glas
Train the segmentation model with MSG modules.
cd MSG
bash tools/dist_train.sh configs/pspnet_mssg/pspnet_wres38-d8_10k_histo.py 4 runs/mssg
Method | mIOU | Weight | Pseudo-mask |
---|---|---|---|
Our MSSG | 62.72% | Download Link (Extracted Code: jjey) | Download Link (Extracted Code: 9zvb) |
SGSCN [1] | 52.61% | Coming Soon | Coming Soon |
PiCIE [2] | 48.77% | Coming Soon | Coming Soon |
*The backbone of all methods above are the same, i.e., PSPNet