Vison307 / PistoSeg

Code Repository for AAAI23 paper "Weakly-Supervised Semantic Segmentation for Histopathology Images Based on Dataset Synthesis and Feature Consistency Constraint"

PistoSeg

Code Repository for AAAI23 Paper "Weakly-Supervised Semantic Segmentation for Histopathology Images Based on Dataset Synthesis and Feature Consistency Constraint"

If you are from China Mainland, you can refer to the gitee repo https://gitee.com/vison307/PistoSeg.

Installation

Code tested on

• Ubuntu 18.04
• A single Nvidia GeForce RTX 3090
• Python 3.8
• Pytorch 1.12.1
• Pytorch Lightning 1.7.1

Please use the following command to install the dependencies:

conda env create -f environment.yaml

Preparing the Data and Weights

1. Download the WSSS4LUAD dataset and put it in ./data/WSSS4LUAD

2. Download the BCSS-WSSS dataset and put it in ./data/BCSS-WSSS (Thanks to Han et. al)

3. Download the ImageNet-pretrained ResNet weight ilsvrc-cls_rna-a1_cls1000_ep-0001.params from SEAM repository and put it in ./weights/ilsvrc-cls_rna-a1_cls1000_ep-0001.params

4. Download the CAM model's weights res38d.pth from OEEM repository and put it in ./weights/res38d.pth

CAM Generation

Please refer to the OEEM readme in ./OEEM/README.md

Or you can use the pre-generated CAMs (in AAAI23/{dataset}/data/CAM/train.zip) and unzip and put them in ./data/WSSS4LUAD/CAM/train and ./data/BCSS-WSSS/CAM/train.

Train & Test

WSSS4LUAD

1. Assure that the CAM of the training set is put into ./data/WSSS4LUAD/CAM/train (in .npy format)

2. Split the validation and test dataset with split_validation.ipynb to produce 224x224 regular-shaped patches

3. Prepare the synthesized dataset and background mask with create_dataset.ipynb

4. Run bash run.sh

BCSS-WSSS

1. Assure that the CAM of the training set is put into ./data/BCSS-WSSS/CAM/train (in .npy format)

2. Prepare the synthesized dataset with create_dataset_bcss.ipynb

3. run bash run-bcss.sh

Note

Due to a disaster to the server, the original weights of the results provided in the paper are lost. So the current codes and weights are re-implemented and re-trained. The results are a bit different from the original ones (but slightly improved in mIoU). The overall performance is still similar.

WSSS4LUAD

• Test mIoU: 0.7530
• Test fwIoU: 0.7582
• Test tissue IoU: [0.7991 0.7020 0.7580] - [TUM, STR, NOM]

BCSS-WSSS

• Test mIoU: 0.7075
• Test fwIoU: 0.7576
• Test tissue IoU: [0.8144 0.7446 0.6063 0.6645] - [TUM, STR, LYM, NEC]

Reproducibility

We tried our best to ensure the reproducibility of the results, but since the torch.nn.functional.interpolate function is not deterministic, the results may be different over runs. If you want to fully reproduce the results, you can use the following weights (code: yj84) (preliminary segmentation epoch=*.ckpt, refining ResNet38-RFM.pth, and precise segmentation segmentation_log/epoch=*.ckpt) and intermediate results (Generated CAM data/CAM/train.zip and Refined Masks refine/CAM.zip). Training logs are also provided for reference.

Citation

If you find our code helpful, please cite as follows:

@inproceedings{fang2023weakly, title={Weakly-supervised semantic segmentation for histopathology images based on dataset synthesis and feature consistency constraint}, author={Fang, Zijie and Chen, Yang and Wang, Yifeng and Wang, Zhi and Ji, Xiangyang and Zhang, Yongbing}, booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, volume={37}, number={1}, pages={606--613}, year={2023} }