MSSG

Qixiang ZHANG, Yi LI, Cheng XUE, Xiaomeng LI*, "Morphology-inspired Unsupervised Gland Segmentation via Selective Semantic Grouping", MICCAI 2023 (Accepted).

1. Introduction

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.

2. Environment

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.

3. Preparation

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.

4. Training for SPM Module

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

5. Training for MSG Module

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

6. Results and Visualization of Pseudo-masks

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