Integrating Quantitative MRI and Artificial Intelligence to Improve Prostate Cancer Classification

Primary outcomes

1. Development of quantitative dynamic contrast (DCE)-enhanced-magnetic resonance imaging (MRI) analysis techniques

   Time frame: Up to 5 years

   Both transfer constant (Ktrans) and rate constant (Kep) from normal prostate tissue will be evaluated for the inter-scanner variability. Pairwise dissimilarities between distributions will be estimated by computing the Kolmogorov-Smirnov statistic, defined as the maximum difference between the empirical distribution functions over the range of the parameter, using 200 cases for each of three MRI scanners. The mean of these pairwise dissimilarities between scanners will be computed to quantify the overall discrepancy of each DCE-MRI model. Construction of a 95% confidence interval for the difference in the mean discrepancies using the nonparametric bootstrap will be done to compare this mean discrepancy between DCE-MRI models. 10,000 bootstrap samples will be generated by sampling patients with replacement, stratifying by the scanner. Will conclude that the proposed DCE-MRI model has a reduced inter-scanner variability if the 95% confidence interval is entirely less than zero.

2. Development of diffusion weighted imaging (DWI) methods that reduce prostate geometric distortion

   Time frame: Up to 5 years

   Differences between rectangular field of view-ENCODE and standard DWI in terms of the prostate Dice's similarity coefficient (primary outcome) and apparent diffusion coefficient consistency will be compared.

3. Development of multi-class deep learning models

   Time frame: Up to 5 years

   The overall performance of FocalNet and Prostate Imaging Reporting & Data System version 2 will be compared in terms of area under the curve. Comparison between area under the curves will be performed using DeLong's test. Will also include the comparison between FocalNet and baseline deep learning methods (U-Net and Deeplab without focal loss [FL] and mutual finding loss [MFL]) to characterize the advantages of using FL and MFL with the same study cohort. For each of these approaches, an optimal cut-point for classification of clinically significant prostate cancer will be identified by maximizing Youden's J (= sensitivity + specificity - 1) and will report sensitivity, specificity and 95% confidence intervals based on the selected cut-point.