FW-RD

Prerequisites

Environment

Python (3.8)
Numpy (1.20.3)
Scipy (1.7.1)
PyTorch (0.3.1)/CUDA 8.0
torchvision (0.2.0)
PIL (8.3.2)
scikit-image (0.13.1)
OpenSlide 3.4.1
matplotlib (2.2.2)
sklearn (1.0)

Dataset

The main data are the whole slide images (WSI) in *.tif format from the Camelyon16 challenge. You may also download the dataset at GigaDB. There are 400 WSis in total, together about 700GB+. Once you download all the slides, please put all the tumor slides and normal slides for training under one same directory.

The Camelyon16 dataset also provides pixel level annotations of tumor regions for each tumor slide in xml format. You can use them to generate tumor masks for each WSI.

Model

This is the main structure of our project. In order to train and test our model, we need to crop WSIs in to 256*256 small patches. The selection and cropping process are shown in Preprocess part.

Preprocess

In order to train WSIs with deep learning models, we need to crop WSIs into 256*256 patches.

Annotations

xml annotation to json

python preprocess/wsi/bin/camelyon16xml2json.py {xml_path} {json_path}

Generate image patches

Generate tissue masks for WSIs

python preprocess/wsi/bin/tissue_mask.py {wsi_path} {tissue_path}

Generate tumor masks for WSIs

python preprocess/wsi/bin/tumor_mask.py {wsi_path} {json_path} {tumor_path}

Generate masks for normal areas

python preprocess/wsi/bin/non_tumor_mask.py {tumor_path} {tissue_path} {normal_path}

Randomly select patches on WSIs on level0

python preprocess/wsi/bin/sampled_spot_gen.py {mask_path} {coords_path} /maximum # of patch for each WSI/

I have included the coordinates of pre-sampled patches used in the paper for training and testing. They are located at perprocess/coords.

Generate patches for training and testing use

python preprocess/wsi/bin/patch_gen.py {wsi_path} {coords_path} {patch_path}

Please save the generated patches with following sequence:

├── data/
│   ├── camelyon16/
│   │   ├── train/
│   │   │   ├── good/
│   │   │   ├── bad/
│   │   ├── test/   
│   │   │   ├── good/
│   │   │   ├── tumor/
│   │   ├── val/   
│   │   │   ├── good/
│   │   │   ├── tumor/

Train and test the model

Run main.py to train and test the model, put the checkpoints in "checkpoints" file.

python main.py

Anomaly Map Generation

set vis = Ture in main.py

python main.py {wsi_path} {ckpt_path} {cfg_path} {mask_path} {probs_map_path}

cfg file for RD model is under '\preprocess\configs'. {mask_path} indicates the tissue mask for test WSIs, you can generate them follow the instruction in Camelyon17 challenge. Please save the anomaly maps for tumor and normal WSIs with following path:

├── results/
│   ├── good/
│   ├── tumor/

Postprocess

Note, Test_049 and Test_114 are excluded from the evaluation as noted by the Camelyon16 organizers.

AUROC Evaluation

AUROC Evaluation

python preprocess/wsi/bin/AUROC_plot.py {probs_map_path}

2.Heatmap Generation

python preprocess/wsi/bin/heatmap_plot.py {prob_map}

FROC Evaluation

Tumor localization We use non-maximal suppression (nms) algorithm to obtain the coordinates of each detectd tumor region at level 0 given a probability map.

python preprocess/wsi/bin/nms.py {probs_map_path} {coord_path}

Where {probs_map_path} is where you saved the generated probability map, and {coord_path} is where you want to save the generated coordinates of each tumor regions at level 0 in csv format. There is an optional command --level with default value 6, and make sure it's consistent with the level used for the corresponding tissue mask and probability map.

2.FROC evaluation With the coordinates of tumor regions for each test WSI, we can finally evaluate the average FROC score of tumor localization.

python preprocess/wsi/bin/Evaluation_FROC.py {Camelyon16_test_image_mask} {coord_path}

{Camelyon16_test_image_mask} is where you put the ground truth tif mask files of the test set, and {coord_path} is where you saved the generated tumor coordinates. Evaluation_FROC.py is based on the evaluation code provided by the Camelyon16 organizers with minor modification.

Reference

@InProceedings{MIDL2023,
author    = {He, Yinsheng and Li, Xingyu},
title     = {Whole-slide-imaging Cancer Metastases Detection and Localization with Limited Tumorous Data}

wollf2008 / FW-RD