Principal Component Analysis of Dynamic Contrast Enhanced MRI in Human Prostate Cancer

Abstract

To develop and evaluate a fast, objective and standardized method for image processing of dynamic contrast enhanced MRI of the prostate based on principal component analysis (PCA). The study was approved by the institutional internal review board; signed informed consent was obtained. MRI of the prostate at 3 Tesla was performed in 21 patients with biopsy proven cancers before radical prostatectomy. Seven 3-dimensional gradient echo datesets, 2 pre and 5 post-gadopentetate dimeglumine injection (0.1 mmol/kg), were acquired within 10.5 minutes at high spatial resolution. PCA of dynamic intensity-scaled (IS) and enhancement-scaled (ES) datasets and analysis by the 3-time points (3TP) method were applied using the latter method for adjusting the PCA eigenvectors. PCA of 7 IS datasets and 6 ES datasets yielded their corresponding eigenvectors and eigenvalues. The first IS-eigenvector captured the major part of the signal variance because of a signal change between the precontrast and the first postcontrast arising from the inhomogeneous surface coil reception profile. The next 2 IS-eigenvectors and the 2 dominant ES-eigenvectors captured signal changes because of tissue contrast-enhancement, whereas the remaining eigenvectors captured noise changes. These eigenvectors were adjusted by rotation to reach congruence with the wash-in and wash-out kinetic parameters defined according to the 3TP method. The IS and ES-eigenvectors and rotation angles were highly reproducible across patients enabling the calculation of a general rotated eigenvector base that served to rapidly and objectively calculate diagnostically relevant projection coefficient maps for new cases. We found for the a priori selected prostate cancer patients that the projection coefficients of the IS-2nd eigenvector provided a higher accuracy for detecting biopsy proven cancers (94% sensitivity, 67% specificity, 80% ppv, and 89% npv) than the projection coefficients of the ES-2nd rotated and non rotated eigenvectors. PCA adjusted to correlate with physiological parameters selects a dominant eigenvector, free of the inhomogeneous radio-frequency field reception-profile and noise-components. Projection coefficient maps of this eigenvector provide a fast, objective, and standardized means for visualizing prostate cancer.