Deep Learning model for prostate cancer diagnosis and cancer stage prediction from H&E histology images

Based on internal benchmarking we settled with DenseNet for the cancer stage prediction, and ResNet56 for the cancer/benign classification task.

For details and medical background check our report at BE_223B_final.docx.pdf.

For details, see report.pdf.

Cancer diagnosis (Cancer vs healthy classification)

Running

python benign_cancerous_resnet56_final.py

performs the following tasks

• create matrices for the images and for the label of the train, val, test sets
• covert to categorical, normalize pixel values, and decide whether to subtract mean
• define a function that creates a DenseNet, and a function that prints the learning rate at the end of each epoch
• perform grid search for hyperparameters
• create a t-SNE plot with PCA preprocessing
• perform manual err analysis
• save plot of confusion matrices at the tile level and the patient level.

Cancer stage (Gleason Level) prediction

Running

python densenet_3class_cancer.py

implements the following pipeline

• create dictionary based on the csv file with the labels
• create matrices for train, test data
• normalize pixel values in the data and split them into train, validation and test sets
• define lr_schedule
• define ResNet layer
• define ResNetv2 architecture
• subtract mean and convert to categorical
• create function evaluate_network which has x_train, x_val as global variables, returns only the val_acc, and is solely used for the Bayesian Hyperparameter Optimization
• create function create_train_return_model, which fulfills a similar goal as evaluate_network but also retunrs the model, to do manual hyperparameter tuning
• implement the Bayesian Hyperparameter Optimization to roughly tune the hyperparameters.
• do some further manual tuning of the Hyperparameters using the function create_train_return_model
• plot training history for best model, and evaluate on test set
• make t-SNE plots
• print statements for manual err analysis
• plot accuracy vs image size

Manual err analysis using grad_cam_heat

grad_cam_heat.ipynb shows what features the model uses to make its prediction in one of the false positive diagnoses.