Gleason_grading

This repository contains the relevant code to utilize a Convolutional Neural Network to classify the Gleason grades of prostate cancer.

The image database is sourced from the Automated Gleason Grading Challenge 2022 (AGGC2022).

To run the model, do the following:

1. copy the github repo: git clone https://github.com/Biomedical-Data-Design-2022-2023/Gleason_grading.git
2. create an environment: conda env create -f environment.yml
3. Go to demo folder python demo.py --i /data/acharl15/gleason_grading/test_folder/Subset1_Test_24.tiff --output_folder ./demo_result/ --model ../neural_network_training/checkpoint/Subset1_epoch36.pth --background "white" --i: define original image path --output_folder: define result path --model: define pretrain model path (saved under ./checkpoint/) --background: define what is the color of the background of the same scans ("white","black").

The output contains:

1. patch_mask.jpg: binary image indicating tissue region
2. G_pred.jpg: Scalar Image, Absolute classification of each patch image. Whiter color means higher grading score. Size of height x weight (where the pixel value is the predicted class label) (0:empty background, 3. 1:normal, etc)
3. G_pred_color.jpg: Colored Heatmap based on G_pred.jpg
4. G_pred_after_mor.jpg: Scalar Image after closing morphological transformation. Size of height x weight (where the pixel value is the predicted class label) (0:empty background, 3. 1:normal, etc)
5. G_pred_after_mor_color.jpg: Colored Heatmap based on G_pred_after_mor.jpg
6. result.pck : saved y-probability, true label, and index in pickel file, which is a dictionary. output[“yprob”] = yprob (size: n*5) output[“ytrue”] = label (size: n) output[“index_x”] = index_list_x (size: n) output[“index_y”] = index_list_y (size: n) Colored image label: Green: normal; Blue: stroma; Yellow: G3; Fuchsia: G4; Red :G5 Each pixel value is corrected by confidence level, where ambiguous color assignment indicates lower confidence level of class assignment.

The automated grading system also contains the following 2 steps:

1. Preprocessing Step: See folder preprocess

2. Neural Network Training: see folder neural_network_training

• Trained model parameters are saved under the neural_network_training/checkpoint folder