Computer-assisted lesion classification and intervention planning for prostate cancer
Abstract
Multi-parametric magnetic resonance imaging (mp-MRI) is emerging as a useful tool for classifying prostate cancer (PCa); however, it suffers from two major limitations: (1) complex, multi-dimensional signals make interpretation challenging and (2) inter-observer variability of lesion classification between physicians. Critically needed are methods for augmenting the interpretability of mp-MRI to assist in lesion classification. To meet this need, we leveraged a patient cohort with post-surgery pathologist-annotated transverse histology images registered to pre-surgery in-vivo mp-MRI with a measured target registration error. We developed a radiomics-based machine learning model trained on annotations for PCa vs. non-PCa, and found that a 5-feature Naïve-Bayes classifier classified these two types of regions in leave-one-patient-out cross-validation with an area under the receiver operating characteristic curve of 0.80. We then investigated augmentation of mp-MRI interpretability via additional imaging using prostate-specific membrane antigen positron emission tomography (PSMA-PET). Using patients imaged on a PET/MRI hybrid-scanner, we investigated segmentation recommendations for classifying the dominant intraprostatic lesion. For focal therapy, we found an optimal threshold of 67% SUVmax and an 8.4 mm margin produced mean voxel-wise sensitivity of 95% with mean specificity 76%. For guided biopsy, we found that a threshold of 81% SUVmax and a 5.2 mm margin produced a mean sensitivity of 65% and mean specificity of 95%. Furthermore, we assessed the effect of boosting high-dose-rate brachytherapy treatment plans towards these segmentations. We found that dose to the underlying high-grade cancer was significantly increased in targeted plans, compared to whole-gland plans, while maintaining all dose constraints.