Presentation Authors: Leo Chen*, Nicholas Bien, Richard Fan, Robin Cheong, Pranav Rajpurkar, Alan Thong, Nancy Wang, Sarir Ahmadi, Mirabela Rusu, James Brooks, Andrew Ng, Geoffrey Sonn, Stanford, CA
Introduction: The role of multiparametric MRI in clinical care is rapidly expanding due to its ability to improve prostate cancer detection. However, MRI interpretation suffers from a false negative rate of around 5-7% and high interobserver variability even among experts, thereby limiting its predictive value. Advances in machine learning and artificial intelligence have the potential to improve and standardize prostate cancer detection on MRI, though studies have relied on radiologist interpretations such as PIRADS scores of segmented lesions as ground truth. We sought to improve over existing methods by directly training on the 3D location and pathology of biopsy cores.
Methods: MR-ultrasound fusion biopsies were performed at a single institution using a robotic fusion biopsy device (Artemis, Eigen) for patients that had multiparametric prostate MRIs. Patients underwent both targeted and standard template biopsies. Core level pathology was prospectively collected into a database. The spatial coordinates of both targeted and standard template cores were calculated and plotted onto the MR images. A weakly supervised convolutional neural network model was trained to predict cancer on MR images, using the spatial geometry and pathology of the biopsy core tracts as ground truth. Implementation was done using the Python programming language.
Results: Over 10,000 MRI-US fusion biopsy cores were collected from over 600 patients in 2015-2018, yielding over 40,000 data points. A preliminary binary classification model based on T2 sequences alone correctly predicted benign versus cancerous cores with an AUROC of 0.78. There is ongoing work on incorporating DWI and ADC sequences to improve the accuracy of the deep learning model.
Conclusions: We present a deep neural network model to predict prostate cancer on MRI that is trained on spatial coordinates and pathology of biopsy cores as ground truth. To our knowledge, this dataset is the largest that has been reported. Our model is blind to radiologist interpretation and lesion segmentation: it thereby aims to eliminate the interobserver variability of radiologist interpretations and improve the detection of cancer in patients with false negative MRIs. Ultimately, our goal is the automated generation of probabilistic heat maps of clinically significant prostate cancer based on automated processing of MRIs, improving the speed and accuracy of detecting clinically significant cancer, while reducing overdetection of clinically insignificant cancer.