Structural and Metabolic Pattern Classification for Detection of Glioblastoma Recurrence and Treatment-Related Effects

Abstract

Artificial neuronal network (ANN) in classification of glioblastoma multiforme (GBM) recurrence from treatment effects using advanced magnetic resonance imaging techniques was evaluated. In 56 patients with treated GBM, normalised minimal and mean apparent-diffusion coefficient (ADC) values, vessels number on susceptibility-weighted images (SWI) and Cho/Cr ratio were analysed statistically and by ANN. Significant correlation exists between normalised minimal and mean ADC values, and no correlation between ADC and Cho/Cr values. Cut-off values for tumour presence were: 1.14 for normalised minimal ADC (54% sensitivity, 71% specificity), 1.13 for normalised mean ADC (51% sensitivity, 71% specificity), 1.8 for Cho/Cr ratio (92% sensitivity, 82% specificity), grade 2 for SWI (87% sensitivity, 82% specificity). An accurate prediction of ANN to classify patients into GBM progression or treatment effects group was 99% during the training and 96.8% during the testing phase. Multi-parametric ANN allows distinction between GBM recurrence and treatment effects, and can be used in clinical practice.