MRSI brain tumor characterization using wavelet and wavelet packets feature spaces and artificial neural networks

Magnetic resonance spectroscopic imaging (MRSI) is a non-invasive technique for assessing biochemical fingerprint of tissue composition. The need to differentiate between normal and abnormal tissues and determine type of abnormality before biopsy or surgery motivated development and application of MRSI. There are several technical reasons that make the brain easier than other organs to be examined with MRSI. This work presents our proposed methods and results for the analysis of the brain spectra of patients with three tumor types (malignant glioma, astrocytoma, and oligodendroglioma). After extracting features from MRSI data using wavelet and wavelet packets, we use artificial neural networks to determine the abnormal spectra and the type of abnormality. We evaluated the proposed methods using clinical and simulated MRSI data and biopsy results. The MRSI analysis results were correct 97% of the time when classifying the spectra of the clinical MRSI data into normal tissue, tumor, and radiation necrosis. They were correct 72% and 83% of the time when determining tumor types using the clinical and simulated MRSI data, respectively.