Machine learning for multiparametric ultrasound classification of prostate cancer using B-mode, shear-wave elastography, and contrast-enhanced ultrasound radiomics

The diagnosis of prostate cancer (PCa) is still based on systematic biopsy, but is increasingly developing towards an imaging-driven approach. In particular, multiparametric magnetic resonance imaging (MRI) is receiving increasing attention over the last few years. In light of MRI-related issues concerning costs, practicality, and availability, we investigate a multiparametric ultrasound (mpUS) approach. We propose and test a machine-learning-based strategy that automatically combines B-mode ultrasound, shear-wave elastography (SWE), and dynamic contrast-enhanced ultrasound (DCE-US) features. To this end, automatic zonal segmentation by deep learning, model-based feature estimation (related to contrast-agent perfusion and dispersion), radiomic feature extraction, and a random-forest-based pixel-wise classification were combined. The method was trained and validated against histopathologically-confirmed benign and malignant regions of interest in 48 PCa patients, in a leave-one-patient-out cross-correlation fashion. The mpUS classification algorithm yielded a region-wise area under the Receiver Operating Characteristics (ROC) curve of 0.75 and 0.90 for PCa and significant (i.e., Gleason ≥4+3) PCa, respectively. In comparison, the best-performing single parameter (i.e., DCE-US-based contrast velocity) reached a performance of 0.69 and 0.76 in terms of the ROC curve area. In particular the combination of perfusion-, dispersion-, and elasticity-related features were favored in the classification. Even though validation on a larger patient group is required, we have demonstrated the technical feasibility of automatic mpUS for PCa localization. Further development of mpUS might lead to a valuable clinical option for robust, ultrasound-driven PCa diagnosis.