Multichannel Texture Analysis Using Localized Spatial Filters
IEEE Transactions on Pattern Analysis and Machine Intelligence
C4.5: programs for machine learning
C4.5: programs for machine learning
Pattern Classification (2nd Edition)
Pattern Classification (2nd Edition)
Data Mining: Practical Machine Learning Tools and Techniques, Second Edition (Morgan Kaufmann Series in Data Management Systems)
Correntropy based feature selection using binary projection
Pattern Recognition
Hi-index | 0.00 |
Although multiparametric (MP) MRI (MP-MRI) is a valuable tool for prostate cancer (CaP) diagnosis, considerable challenges remain in the ability to quantitatively combine different MRI parameters to train integrated, fused meta-classifiers for in vivo disease detection and characterization. To deal with the large number of MRI parameters, dimensionality reduction schemes such as principal component analysis (PCA) are needed to embed the data into a reduced subspace to facilitate classifier building. However, while features in the embedding space do not provide physical interpretability, direct feature selection in the high-dimensional space is encumbered by the curse of dimensionality. The goal of this work is to identify the most discriminating MP-MRI features for CaP diagnosis and grading based on their contributions in the reduced embedding obtained by performing PCA on the full MP-MRI feature space. In this work we demonstrate that a scheme called variable importance projection (VIP) can be employed in conjunction with PCA to identify the most discriminatory attributes. We apply our new PCA-VIP scheme to discover MP-MRI markers for discrimination between (a) CaP and benign tissue using 12 studies comprised of T2-w, DWI, and DCE MRI protocols and (b) high and low grade CaP using 36 MRS studies. The PCA-VIP score identified ADC values obtained from Diffusion and Gabor gradient texture features extracted from T2- w MRI as being most significant for CaP diagnosis. Our method also identified 3 metabolites that play a role in CaP detection--polyamine, citrate, and choline--and 4 metabolites that differentially express in low and high grade CaP: citrate, choline, polyamine, and creatine. The PCA-VIP scheme offers an alternative to traditional feature selection schemes that are encumbered by the curse of dimensionality.