Supervised pattern recognition for predicting the appearance of contrast-enhancement in high-grade gliomas using multimodality MRI and MRSI
Chang S, Nelson S, Lee M
University of California
The appearance of contrast-enhancement in high-grade gliomas indicates a significant breakdown of the blood-brain barrier. We are interested in determining which spectroscopic, diffusion, and perfusion parameters are most predictive of the appearance of contrast enhancement at a later date. We hypothesize that data obtained prior to radiotherapy provide information that can be useful in predicting the location of the contrast-enhancing lesion after radiotherapy. A genetic algorithm is used to select and scale features for an optimized k-nearest-neighbor classification algorithm. By comparing the receiver-operating characteristic curves, we demonstrate that the optimized classifier significantly outperforms thresholding on any single-variable in the study.