AI grading

This repository is for the paper "The artificial intelligence-based model ANORAK improves histopathological grading of lung adenocarcinoma" published in Nature Cancer. It could guide you to generate growth pattern masks with a well-trained deep learning model for semantic segmentation, from which the proportion of each growth pattern can be obtained, thereby replicating IASLC grading for lung adenocarcinoma.

Generating tiles for whole slide images

Dependencies for generating_tiles are in AIgrading/generating_tiles/requirements, following the step in the work of AbdulJabbar, K. et al. Geospatial immune variability illuminates differential evolution of lung adenocarcinoma. Nature Medicine (2020). doi: 10.1038/s41591-020-0900-x

After all dependencies are well installed, if the format of whole slide images are in .svs format, then

python ./generating_tiles/main_tiles.py -d /path/to/raw/slides -o /path/to/result -p '*.svs'

The ouptut structure will be

result_cws_tiling/
        ├── TCGA-xxxx-xxxx.svs
		├── Da0.jpg
		├── Da1.jpg
		└── ...
	├── TCGA-xxxx-xxxx.svs
		├── Da0.jpg
		├── Da1.jpg
		└── ...

Training

Dependencies for training_patch are in AIgrading/requirements.

Dataset for training: https://doi.org/10.5281/zenodo.10016027

The trained model checkpoint is deposited in https://zenodo.org/records/15272883

Step0: creat the conda environment following AIgrading/requirements.txt

Step1: divide image in trainset (download from 10.5281/zenodo.10016027) into patches with a size of $768 \times 768$

python ./training_patch/img2patch.py --image_path /path/to/training/image --label_path /path/to/training/mask --save_path /path/to/patches

The output training patches from this step are structured as

Training_patches /
├── image
│   ├── train001_xxx_0.png
│   ├── train001_xxx_1.png
│   └── ...
└── maskPng
    ├── train001_xxx_0.png
    ├── train001_xxx_1.png
    └── ...

Step2: train the model

python ./training_patch/train_main.py --input_dir /path/to/patches/image --target_dir /path/to/patches/maskPng --img_size 384 --num_class 7 --batch_size 8 --num_epoch 60

Inference

Dependencies for inference_slide are in AIgraind/requirements, same with the training

The trained model checkpoint is deposited in https://zenodo.org/records/15272883

Input: H&E image tiles

Output: Growth pattern mask,

#000000 black-background, 
#0000ff blue-lepidic, 
#ffff00 yellow-papillary, 
#ff0000 red-acinar, 
#00ffff cyan-cribriform, 
#ff00ff magenta-micropapillary, 
#880000 dark red-solid

python ./inference_slide/main_gp.py -d /path/to/cws_tiling -o /path/to/cws_mask -s /path/to/ss1_mask -sf /path/to/ss1_mask_final -p *.svs -n 0