Skip to content

Latest commit

 

History

History
65 lines (51 loc) · 2.69 KB

README.md

File metadata and controls

65 lines (51 loc) · 2.69 KB
Raw

Label-Effcient-Nuclei-Segmentation

Official code implementation for paper 'Which Pixel to Annotate: a Label-Efficient Nuclei Segmentation Framework'

The whole framework consists of 3 stages: 1. Patch selection 2. CSinGAN data augmentation 3. Pseudo-label based MRCNN

Stage 1: Consistency-based Patch Selection

In this stage, patches with representativity and consistency will be selected for data augmentation and semi-supervised segmentation.

Data

The dataset that we used in this paper is the Kumar Monusag; TNBC ;

In first-stage clustering, the images are cropped into 256*256 patches and stored in 'data/'.

In second-stage clustering, the 256*256 patches are cropped into 128*128 patches and stored in 'data_cut'.

The selected results are the folder names in 'data/'. The tunable parameters are K1, K2 which are the clutser numbers for 2 stages.

Code

python consis_based.py

Stage 2: CSinGAN augmentation

In stage 2, the random masks are synthesized by randomly permuting the nuclei positions in the selected patches' mask. Then the real pair and random masks are the inputs for CSinGAN.

Training Data

train.png: the selected real patch

train_mask.png: the corresponding mask for the real patch

randon mask/: the synthesized masks that we want to add patterns

Code

Dependencies

torch==1.4.0; torchvision==0.5.0 ; opencv

Mask synthesis

python mask_synthesis.py

Training

python main_train.py

Stage 3: Pseudo-label based MRCNN

This Mask-rcnn code implementation is modified by Matterport's implementation of Mask-RCNN: https://github.com/matterport/Mask_RCNN. The evaluation code is modified for pseudo label prediction. The environment dependencies and data strcture are following the same setting with Matterport's codes.

Pseudo-label.py

This code is to predict pseudo label for cropped image patches to perform semi-supervised learning. The pretrained model is trained by the synthesis images created by CSinGAN. And the unlabelled images with predicted pseudo labels can form new image paires for new MRCNN training.

Citation

If you find this research useful in your work, please acknowledge it appropriately and cite the paper:

@article{lou2022pixel,
  title={Which Pixel to Annotate: a Label-Efficient Nuclei Segmentation Framework},
  author={Lou, Wei and Li, Haofeng and Li, Guanbin and Han, Xiaoguang and Wan, Xiang},
  journal={IEEE Transactions on Medical Imaging},
  year={2022},
  publisher={IEEE}
}