Deep Learning with Bayesian Inference for Prostate Cancer Diagnosis across Longitudinal Biparametric MRI

BACKGROUND: Despite increasing use of active surveillance via biparametric MR imaging (bpMRI) for prostate cancer (PCa) management, there is a lack of research in medical image computing that utilize longitudinal studies to assist present-day diagnosis. PURPOSE: To investigate the efficacy of a deep learning-based PCa detection model, that integrates past bpMRI exams and population-level ana-tomical priors via Bayesian inference. MATERIALS AND METHODS: This retrospective study included 250 consecutive biopsy-naive men (median age: 64 yrs; IQR: 60-69) with elevated levels of PSA (me-dian level: 8 ng/mL; IQR: 5-11), who underwent at least two consecutive MRI exams between 2016-2018 (N=500). Intra-patient bpMRI scans were rigidly registered, paired with expert voxel-level annotations of PI-RADS 2-5 lesions and subsequently used to train a deep learning model to predict and localize all PI-RADS findings. Radiologists utilize prior studies to inform present-day diagnosis. Similar-ly, for each patient case, computer-aided diagnosis for the follow-up bpMRI exam was derived via Bayesian modelling; probabilistical-ly integrating past bpMRI exams and a population prior for spatial PCa prevalence and zonal anatomy. Diagnostic performance was evaluated by the ability to accurately discriminate patients with benign prostatic tissue (n=10) or PI-RADS <= 3 lesions (n=159), from those carrying PI-RADS >= 4 lesions (n=81), over 5-fold cross-validation. Normalized Wilcoxon Mann-Whitney U statistic was used to derive AUROC and confidence intervals were computed over 5000 replications of bootstrapping. RESULTS: Computer-aided diagnosis of follow-up studies without priori, yielded an AUROC of 0.77 (95% CI: 0.71, 0.83), F0.5 score of 0.51 (95% CI: 0.42, 0.61), positive predictive value (PPV) of 0.51 (95% CI: 0.40, 0.60) and negative predictive value (NPV) of 0.78 (95% CI: 0.71, 0.83). Computer-aided diagnosis of follow-up studies with the inclusion of priori via Bayesian inference yielded an AUROC of 0.80 (95% CI: 0.76, 0.84), F0.5 score of 0.58 (95% CI: 0.48, 0.68), PPV of 0.60 (95% CI: 0.48, 0.72) and NPV of 0.78 (95% CI: 0.72, 0.85).In comparison to stand-alone diagnosis, factoring in prior studies resulted in a 41.4% reduction of AUROC standard deviation across each fold. CONCLUSION: Incorporating past studies and clinical priors via Bayesian inference can improve diagnostic certainty and robustness of deep learning in follow-up patient exams. CLINICAL RELEVANCE/APPLICATION: Prostate cancer is one of the most prevalent cancers in men worldwide. In the absence of experienced radiologists, its morphological heterogeneity can lead to low inter-reader agreement. Automated, reliable detection algorithms can improve diagnostic accuracy with consistent quantitative analysis.