Prostate cancer characterization with a data-driven analysis of multi-center 3D 1H-MRSI data using Multivariate Curve Resolution

Location and aggressiveness of cancer lesions in the prostate were assessed with a Multivariate Curve Resolution (MCR) approach. This data-driven method was trained on 3D 1H MRSI data from 63 patients from 6 institutions to extract common spectroscopic components without needing prior knowledge. MCR identified 5 components related to known prostate metabolites, residual lipid and water signals. The relative amounts of these components robustly separated cancer from benign tissue in a test set of 43 other subjects from 8 centers, and showed a strong correlation with cancer aggressiveness. This approach facilitates automatic analysis of prostate MRSI, avoiding subjective human intervention.