Novel PCA-VIP scheme for ranking MRI protocols and identifying computer-extracted MRI measurements associated with central gland and peripheral zone prostate tumors

• Published in: Journal of magnetic resonance imaging Vol. 41; no. 5; pp. 1383 - 1393
• Main Authors: Ginsburg, Shoshana B., Viswanath, Satish E., Bloch, B. Nicolas, Rofsky, Neil M., Genega, Elizabeth M., Lenkinski, Robert E., Madabhushi, Anant
• Format: Journal Article
• Language: English
• Published: United States Blackwell Publishing Ltd 01.05.2015; Wiley Subscription Services, Inc
• Subjects: Aged; Algorithms; computer-extracted features; Data Interpretation, Statistical; Diffusion Magnetic Resonance Imaging - methods; feature selection; Humans; Image Enhancement - methods; Image Interpretation, Computer-Assisted - methods; Machine Learning; Magnetic resonance imaging; Male; Middle Aged; model interpretation; Pattern Recognition, Automated - methods; Principal Component Analysis; prostate cancer; Prostatic Neoplasms - pathology; Reproducibility of Results; Sensitivity and Specificity; model interpretation; prostate cancer; computer-extracted features; feature selection; principal component analysis
• ISSN: 1053-1807; 1522-2586; 1522-2586
• DOI: 10.1002/jmri.24676

Purpose To identify computer‐extracted features for central gland and peripheral zone prostate cancer localization on multiparametric magnetic resonance imaging (MRI). Materials and Methods Preoperative T2‐weighted (T2w), diffusion‐weighted imaging (DWI), and dynamic contrast‐enhanced (DCE) MRI were acquired from 23 men with confirmed prostate cancer. Following radical prostatectomy, the cancer extent was delineated by a pathologist on ex vivo histology and mapped to MRI by nonlinear registration of histology and corresponding MRI slices. In all, 244 computer‐extracted features were extracted from MRI, and principal component analysis (PCA) was employed to reduce the data dimensionality so that a generalizable classifier could be constructed. A novel variable importance on projection (VIP) measure for PCA (PCA‐VIP) was leveraged to identify computer‐extracted MRI features that discriminate between cancer and normal prostate, and these features were used to construct classifiers for cancer localization. Results Classifiers using features selected by PCA‐VIP yielded an area under the curve (AUC) of 0.79 and 0.85 for peripheral zone and central gland tumors, respectively. For tumor localization in the central gland, T2w, DCE, and DWI MRI features contributed 71.6%, 18.1%, and 10.2%, respectively; for peripheral zone tumors T2w, DCE, and DWI MRI contributed 29.6%, 21.7%, and 48.7%, respectively. Conclusion PCA‐VIP identified relatively stable subsets of MRI features that performed well in localizing prostate cancer on MRI. J. Magn. Reson. Imaging 2015;41:1383–1393. © 2014 Wiley Periodicals, Inc.