Alan Priester1,2, Richard Fan3, Joshua Shubert2, Jonhas Colina2, Mirabela Rusu3, Sulaiman Vesal3,
Wei Shao3, Yash Samir Khandwala3, Shyam Natarajan1,2, Geoffrey A. Sonn3
1 University of California, Los Angeles
2 Avenda Health, Inc.
3 Stanford University
Introduction: A machine learning (ML) algorithm was developed to estimate voxel-level risk of clinically significant prostate cancer (csPCa), resulting in a 3D lesion heat map (LHM). Treatment margins created by thresholding the LHM were retrospectively assessed using whole mount (WM) prostatectomy data as ground-truth. ML margins were compared to standard of care (SOC) methodology, i.e. hemi-gland margins or a 10-mm isotropic expansion of MRI-visible regions of interest (ROIs).
Methods: A machine learning model was developed using multi-institutional data from 875 patients. Input data consisted of T2-weighted MRI, surface models of the prostate, ROIs defined using PI-RADS v2, and tracked biopsy cores (Fig A-B). The model combined a convolutional neural network with a gradient-boosted decision tree, and was trained using 5-fold cross validation. WM data from an external institution (N = 50, Stanford University) was used to evaluate the LHM. All test cases bore MRI-visible, biopsy-confirmed GG2-3 disease apparently isolated to a single hemisphere or the anterior gland. LHMs were generated for each case (Fig C), from which the ML algorithm selected a default margin (Fig E) intended to maximize csPCa encapsulation while limiting margin volume. SOC margins were likewise generated (Fig F). Using WM (Fig D) as ground truth, csPCa-bearing voxels were compared to ML and SOC margins using sensitivity, specificity, and complete csPCa encapsulation rate.
Results: ML margins significantly improved sensitivity for csPCa-bearing voxels (97% versus 94%, p < 0.001, Wilcoxon signed-rank test) and the per-patient rate of complete csPCa encapsulation (80% versus 56%, p = 0.01, chi squared test) relative to hemi-gland margins. ML margins had higher mean sensitivity and csPCa encapsulation than 10-mm margins, though the differences fell short of statistical significance.
Conclusion: A ML model produced margins that were superior to hemi-gland margins, dramatically improving rates of complete csPCa encapsulation without significantly reducing specificity. This treatment planning approach may improve outcomes in focal therapy and warrants further study.
