Automatic Detection of Lipid Peaks in MR Spectroscopic Image Data using Artificial Neural Networks

Introduction: Presence of lipids in brain parenchyma may indicate the active deor re-myelination in multiple sclerosis (MS). Magnetic resonance spectroscopic imaging (MRSI) can provide tissue biochemical information in vivo. Manual examination of each voxel for the presence of lipid peak in MRSI data is tedious due to the number of voxels. This problem becomes more acute in multicenter clinical trials where large data need to be analyzed. To alleviate this problem, we have developed an automatic and rapid procedure that based on artificial neural networks (ANN) for identification of lipid peaks in MRSI data. Methods: Data Acquisition: Proton MRSI data was acquired on 1.5 T scanner with a spin echo sequence with variable TR (maximum of 1000 ms) and TE equals to 30 ms. Acquisition was localized to centrum semiovale region in brain with approximately 100 mm (A/P) x 100 mm (R/L) x 15 mm (S/I) volume. Three chemical shift selective (CHESS) pulses for water suppression and eight outer volume suppression pulses for minimizing extramenengeal tissue contamination were incorporated into the sequence. Other acquisition parameters were: spectral bandwidth=1000 Hz, number of complex points=256, FOV=240x240 mm, and number of phase-encoding steps=32x32. In addition, water unsuppressed MRSI data were also acquired with identical parameters as the metabolite data, except for 16x16 phase-encoding steps, for automatic spectral processing [1]. Localizer image was acquired to generate a mask of the spectroscopic volume-of-interest (VOI). Spectral preprocessing and phase correction were performed as suggested in ref [1]. RBFNN: The radial basis function neural network (RBFNN) can be represented by the parametric model [2]: