An MRI-Based Convolutional Neural Network to Predict Biochemical Recurrence Following Radiotherapy for Intermediate and High-Risk Localized Prostate Cancer.

Abstract

The risk of biochemical recurrence (BCR) following radiotherapy (RT) for localized prostate cancer (LPCa) varies considerably within risk stratification groups defined by classic clinical and pathologic variables; there is an unmet need for low-cost tools that more robustly predict BCR and allow for individualized therapy. Published imaging-based algorithms for BCR prediction after RT are limited to hand-crafted radiomics and/or small cohorts. We aimed to develop a deep learning model to predict BCR at 5 years after RT for intermediate and high risk LPCa using pre-treatment T2-weighted (T2W) MRI. Patients with intermediate and high risk LPCa treated with radical RT at our institution between 2010 and 2015 were included. We excluded those who did not have a pre-treatment T2W-MRI and those with less than 5 years of follow-up. The Phoenix definition for BCR was used. The dataset (DS1) was split into training (70%), validation (20%), and test (10%) sets using a stratified technique. A U-Net model for prostate segmentation was trained and tested on a separate annotated prostate T2W-MRI dataset (DS2) of 225 patients from our institution. The U-Net model was then used to segment the whole prostate gland on the MRI images of DS1, and the segmented images were fed into four 2D convolutional neural networks (CNNs) using different network architectures and regularization techniques (VGG blocks with batch normalization, dropout, and max pooling layers) to predict BCR at 5 years. The CNNs were evaluated using the area under the receiver operating characteristic curve (AUC) on the test set. For benchmarking, three machine learning classifiers (Random Forest, Logistic Regression, and Support Vector Machines) were developed using the 5 most important features selected by Mean Decrease in Impurity from a set of 18 clinical variables. A total of 189 patients were included in DS1. Androgen deprivation therapy (ADT) was received by 83.6% of patients. BCR was identified in 26% of the cases. The Dice score for the U-Net segmentation model was 78% on the test set of DS2. The AUC achieved by the different CNNs for predicting BCR ranged between 0.53 and 0.75. The best performing CNN consisted of 3 convolutional layers, the first two followed by max-pooling layers, a flattening layer, a dense layer, and an output layer with softmax activation function. The best clinical model was a Random Forest algorithm with an AUC of 0.70. The selected clinical variables by decreasing feature importance were: age, time to nadir PSA, pre-treatment PSA, percentage of positive biopsy cores at diagnosis, and nadir PSA. We developed a deep learning model based on pre-treatment T2W-MRI to predict BCR at 5 years following radical RT for intermediate and high-risk LPCa. This CNN outperformed a model based on clinical variables and warrants further validation in external cohorts.