This repository contains scripts to reproduce the analysis presented in the paper "Automated Gleason grading of prostate cancer tissue microarrays via deep learning". The corresponding tissue microarray dataset is available on Harvard Dataverse.
The first step is to create image patches for training and validation (utils/create_patches.py). We then train a convolutional neural network, e.g. using the MobileNet architecture, to classify image patches as benign, Gleason pattern 3, 4 or 5. Given the small size of our dataset, fine-tuning models pre-trained on ImageNet achieved better results (gleason_score_finetune.py).
Once having a good model (the model used in the manuscript is provided: model_weights/MobileNet_gleason_weights.h5), we can use it to produce pixel-level probability maps for entire TMA spots and visualise class-discriminative regions via class activation mappings (plot_heatmaps_and_CAM.py). We only want to make predictions on tissue regions, so we first need to create tissue region masks for the TMA images (utils/create_tissue_masks.py). Finally, we compare the model's and pathologists' Gleason annotations on the test cohort (plot_test_cohort_results.ipynb).
Our scripts assume the following dataset structure:
/path/to/dataset_TMA
|-- TMA_images # All H&E stained images (TMAs 76, 80, 111, 199, 204)
|-- Gleason_masks_train # Gleason annotations for training and validation (TMAs 76, 111, 199, 204)
|-- Gleason_masks_test
|-- Gleason_masks_test_pathologist1 # Test set annotations by first pathologist (TMA 80)
|-- Gleason_masks_test_pathologist2 # Test set annotations by secind pathologist (TMA 80)