MP74-10 DEEP LEARNING FOR SEMI-AUTOMATED PIRADSV2 SCORING ON MULTIPARAMETRIC PROSTATE MRI

Abstract

INTRODUCTION AND OBJECTIVES: Multiparametric MRI (mpMRI) of the prostate has demonstrated great promise in aiding diagnosis of clinically significant prostate cancer. The Prostate Imaging Reporting and Data System (PIRADS)v2 was designed to provide standardized interpretation and reporting but inter-reader agreement remains low. In this study, we use a deep learning (DL) approach to develop a semi-automated PIRADSv2 scoring system. METHODS: Patients undergoing mpMRI for assessment of localized disease from February - October 2018 were considered for study inclusion. mpMRI scans were acquired at 3T with T2-weighted (T2W), diffusion-weighted (ADC map, high b-value DWI acquisition), and dynamic-contrast enhanced (DCE) MRI sequences and were prospectively interpreted by a single expert reader with >10years experience. All detected lesions were manually contoured on T2W MRI with reference to full mpMRI series to capture full extent of lesion boundaries. Bounding boxes derived from lesion contours were fed through a convolutional neural network with a ResNet101 backbone from the fast.ai library as three-channel images (T2W, ADC, high-b). Binary image classification was evaluated comparing PIRADS 2-3 versus PIRADS 4-5. Model performance was further assessed when training with all mpMRI slices encompassing lesion volumes or single mpMRI slice containing largest diameter of each lesion. Data were partitioned to 80% training and 20% validation. Outcome was measured as percent correct classification on the validation dataset. RESULTS: In total, 196 patients with contoured lesions were included in preliminary cohort had a PIRADS score of 2 or higher (n=24 PIRADS 2, n=73 PIRADS 3, n=58 PIRADS 4, n=41 PIRADS 5). When each slice of the tumors were treated as a unique sample, there were a total of 444 segmented tumor slices: n=33 PIRADS 2, n=113 PIRADS 3, n=108 PIRADS 4, n=190 PIRADS 5. Correct classification rate was 60% when the largest slice for each tumor was used. Correct classification rate increased to 71% when all 444 tumor slices were utilized. The PPV, NPV, sensitivity, specificity for correct depiction of PIRADS 4/5 lesions were 87%, 38%, 74%, 58%, respectively. CONCLUSIONS: We have demonstrated the possibility of more accurate and consistent classification of PIRADSv2 scores of lesions detected at mpMRI using an in house-made DL based algorithm. The accuracy of the system showed improvement as the size of the dataset increased in size. Future attempts will focus on further increasing of the dataset size and optimization of the system.