An application of neural networks and ¹H MR in vivo spectroscopy in brain tumors recognition

As the brain is inaccessible for direct examination, imaging techniques are the most important diagnostic tools in the management of patients with primary brain tumors. Both MR and CT images allow for the visualization of brain structures and abnormalities but they lack functional information. The enhancing lesion as seen on CT or MRI not always corresponds to viable tumor especially after the surgery or radiotherapy. Proton MR spectroscopy offers an in vivo non invasive method to assess metabolic changes in brain tissue. The employment of this method as a diagnostic tool provides additional information and helps to diagnose correctly.
MRI localized ¹H MR spectra were acquired from the volumes of interest of 1.5x1.5x1.5 cm³ using a single voxel double-spin-echo PRESS sequence with TR=1 500 ms, TE=35 ms and 100 Acq. In order to reduce any potential sources of errors the spectra were resolved using the automated fitting in the frequency domain with the second derivative method. The obtained data – the main metabolites ratios as well as the normalized integral intensities – were classified and interpreted using the Artificial Neural Networks. 48 spectra acquired from the astrocytoma anaplasticum (AA) tumor beds and 24 ones obtained from the glioblastoma multiforme (GBM) tumor beds were analysed. Normal data were obtained from 18 healthy volunteers.
The aim of the study was evaluation of the decision making process of neural network and to prepare learning data sets for the classification process. Classification abilities of two artificial neural networks were compared. The first one was learned by error back-propagation algorithm, and the other by Simplified Fuzzy Adaptive Resonance Map (SFAM) algorithm.
Our results show that even simple learning algorithms can effectively classify tumor and normal brain tissue. However, in order to distinguish between AA and GBM it is necessary to apply advanced algorithms. In the case of the error back-propagation method the accuracy is 81% and the sensitivity equals 87% whereas for SFAM the appropriate values are 90 and 93%.
The study provides preliminary results on analyzing in vivo MR spectra using Artificial Intelligence Systems.