Radio-pathomic maps of histo-morphometric features trained with whole mount prostate histology distinguish prostate cancer on MP-MRI
Authors
Affiliations (1)
Affiliations (1)
- Medical College of Wisconsin
Abstract
BackgroundProstate cancer (PCa) is the most prevalent male cancer in the U.S., accounting for 29% of new cancer diagnoses. Multiparametric MRI (MP-MRI), including T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps, is an effective tool for detecting PCa; however, accuracy varies, and false-positives may lead to unnecessary biopsies or overtreatment. Radio-pathomic maps (RPMs), derived from MP-MRI and machine learning, have been advantageous in differentiating clinically significant PCa. This study tested whether RPMs of tissue density and histo-morphometric features could better predict cancer presence than conventional MR imaging. Materials and MethodsMP-MRI from 236 patients prospectively recruited between 2014 and 2023 with confirmed PCa were analyzed. Whole-mount prostate sections sliced to match the MRI were processed, digitized, and Gleason-pattern annotated by a GU pathologist. Automated algorithms identified glands and calculated quantitative histo-morphometric features, which were mapped across whole slide images. Slides were nonlinearly aligned to each patients T2WI using in-house software, enabling direct comparison of slides, features, and annotations in MR-space. A multi-step prediction model was trained using a 2/3 - 1/3 train/test split to predict histo-morphometric features using 5x5 voxel tiles from T2WI and ADC. These feature maps were then used generate tumor probability maps. ResultsHistological feature models produced RMSE values approximately within one standard deviation of the ground truths variability, indicating acceptable performance. The best RPM, using histological density features, achieved an accuracy of [~]80%. Visual inspection of RPMs showed good concordance to high-grade cancer annotations. ConclusionThis study demonstrates that the use of MRI intensities can predict complex histo-morphometric features and delineate regions of PCa non-invasively. Future research is warranted to determine the clinical benefit of using RPMs in treatment guidance.