Gliomas classification by multivariate analysis of in vivo MRI/MRSI data based on recursive partitioning tree and discriminant analysis

Accurate diagnosis is critical for the treatment planning of brain tumors. At present, classification of tumor is based on histological examination of tissue samples. This is invasive and may be subject to sampling errors. Magnetic resonance spectroscopic imaging (MRSI) is a non-invasive technique that provides functional information and has been proposed as a tool for non-invasive tumor grading. The goal of this work is to find a classification method that (1) explicitly combines information from MRSI and MR imaging (MRI); (2) considers the MRSI characteristics of the entire lesion instead of a pre-selected region from within the anatomic lesion. Forty-nine newly-diagnosed glioma patients were studied with multivariate analysis based on recursive partitioning analysis (RPA) and linear discriminant analysis (LDA). The cross-validation classification error was 5 out of 49 patients. This suggested that characterizing the lesion by integrating the MRI/MRSI properties has the potential for improving the diagnosis and management of brain tumors.