Prostate-cancer imaging using machine-learning classifiers: Potential value for guiding biopsies, targeting therapy, and monitoring treatment

Abstract

Prostate cancer (PCa) remains a major health concern in many countries. However, it cannot be imaged reliably by any commonly used imaging modality. Therefore, needle biopsies and treatments cannot be targeted to suspicious regions. Our objective is to develop and test an ultrasonic method based on spectrum analysis of radio-frequency (RF) ultrasound echo signals and on classification using current machine-learning tools for reliably imaging PCa and thereby guiding biopsies, targeting therapy, and eventually, monitoring treatment of PCa. RF data were acquired in the biopsy plane of 617 prostate biopsy cores obtained from 64 suspected prostate-cancer (PCa) patients. For each patient, clinical data such as PSA level also were recorded. A level of suspicion (LOS) was assigned based on the conventional image. Spectral computations were performed on acquired RF data in a region of interest that spatially matched the tissue-sampling location. Four non-linear classifiers were trained from these data using biopsy results as the gold standard: multi-layer-perceptron artificial neural networks (ANNs), logitboost algorithms (LBAs), support-vector machines (SVMs), and stacked, restricted Boltzmann machines (S-RBMs). Cross-validation methods were employed to obtain tissue-category scores. Areas under ROC curves (AUCs) were used to assess classifier performance in comparison with LOS-based performance. AUCs for the ANN, LBA, SVM, and RBM respectively were 0.84 ± 0.02, 0.87 ± 0.04, 0.89 ± 0.04, and 0.91 ± 0.04. In comparison, the LOS-based AUC was 0.64 ± 0.03. Tissue-type images (TTIs) based on these methods revealed cancerous foci that subsequently were identified histologically, but were undetected prior to prostatectomy pathology. The ultrasonic imaging methods described here show significant potential for achieving needed reliability. A clinically significant beneficial reduction in false-negative biopsy procedures would be possible if TTIs were used to guide biopsies. Benefits also would result from using TTIs to target focal treatment and reduce toxic side effects. Potentially, TTIs also could be used to assess tissue changes over time for active surveillance and therapy monitoring.