Artificial intelligence-based histopathology image analysis identifies a novel subset of endometrial cancers with distinct genomic features and unfavourable outcome

Development Information

Date Created: 16 April 2024
Developer: Amirali Darbandsari
Version: 0.0

About The Project

The following GIF depicts the proposed workflow in our study. This repo implements the AI step of the workflow.

Installation

mkdir AttentionMIL
cd AttentionMIL
git clone git clone https://github.com/AIMLab-UBC/EC-p53abnlike-AIclassifier .
pip install -r requirements.txt

Usage

From a high-level perspective, AttentionMIL can be divided into three stages:

1. Deriving embeddings from patches
2. Training/Evaluating the network

Each subsection below includes sample configurations.

Deriving the embeddings

Following script, provide sample settings to calculate embeddings the extracted patches:

python3 run.py --experiment_name exp_name \
--log_dir path_to_dir \
--chunk_file_location path_to_json_file \
--patch_pattern pattern_of_patches \
--subtypes subtypes \
--num_classes nb_subtypes \
--backbone resnet34 \
calculate-representation \
--method Vanilla \
--saved_model_location path_to_trained_network \

The above configuration generates the embeddings of the extracted patches using the trained network whose weights located at saved_model_location and writes them in the directory located at path_to_dir/exp_name/representation as pickle files.

Training the network

Following script, provide sample settings to train the AttentionMIL:

python3 run.py --experiment_name exp_name \
--log_dir path_to_dir \
--chunk_file_location path_to_json_file \
--patch_pattern pattern_of_patches \
--subtypes subtypes \
--num_classes nb_subtypes \
--backbone resnet34 \
train-attention \
--lr 0.0001 \
--wd 0.00001 \
--epochs 30 \
--optimizer Adam \
--patience 10 \
--lr_patience 5 \
--use_schedular \
VarMIL

Evaluating the network

Following script, provide sample settings to test the trained network:

python3 run.py --experiment_name exp_name \
--log_dir path_to_dir \
--chunk_file_location path_to_json_file \
--patch_pattern pattern_of_patches \
--subtypes subtypes \
--num_classes nb_subtypes \
--backbone resnet34 \
train-attention \
--only_test \
VarMIL

The network generates a .pkl file at path_to_dir/exp_name/information/VarMIL consisting of predictions, attention mappings, and evaluation metrics such as AUC, accuracy, and F1-score.

Use AttentionMIL on your data

To run AttentionMIL on your own data, you simply need to generate a json file containing the path of extracted patches.

Sample JSON

Each file consists of three IDs (0, 1, and 2), with 0 representing training data, 1 representing validation data, and 2 representing test data.

{"chunks": [{"id": 0, "imgs": ["pattern_of_patches/x1_y1.png", "pattern_of_patches/x2_y2.png"]}, {"id": 1, "imgs": ["pattern_of_patches/x3_y3.png", "pattern_of_patches/x4_y4.png"]}, {"id": 2, "imgs": ["pattern_of_patches/x5_y5.png", "pattern_of_patches/x6_y6.png"]}]}