Quantification and Classification of MR spectra from human brain 量化的MR光谱从人类大脑的分类.pptVIP

Quantification and Classification of MR spectra from human brain tumors using metabolite concentrations as input Overview Quantification of long echo proton MRS signals Prior knowledge Strategy: spectral quantification using MRUI Classification of brain tumors Approaches to classification LS-SVM as a classification algorithm Classification based on the raw spectra data magnitude spectrum range 4.17 ? 0.94 ppm (or -34 ? -240 Hz): 107 points scaling of the data regularisation: determine LS-SVM hyperparameters s, g classification with LS-SVM Classification based on estimated amplitudes Classification based on estimated amplitudes Classification based on estimated amplitudes Discussion points Clinical information available? To what extent do the spectra belong to the assigned classes? e.g. tumor assigned as glioblastomas, while a large part of it is necrosis? information available? is this visible on MR images? how to handle this additional information? Discussion points Combination of data from different centra? introduce additional categorical variable into classification method to correct for differences in centre e.g. CDP 1 SGHMS 3 IDI 2 Nijmegen 4 … do not scale these categorical variables Andy Devos, Leentje Vanhamme, Sabine Van Huffel, Lukas Lukas Department of Electrical Engineering KULeuven Quantification of long echo proton MRS signals Classification of brain tumors Discussion points PR Nijmegen 2001 ESAT-SISTA/COSIC, KULeuven The dataset used Modeling a MRS signal Prior knowledge Strategy: spectral quantification using MRUI PR Nijmegen 2001 ESAT-SISTA/COSIC, KULeuven The dataset used Centre: IDI, Barcelona (Philips data) 2 types of tumors: Glioblastomas= group 1 Meningiomas= group 2 20 data of group 1 23 data of group 2 PR Nijmegen 2001 ESAT-SISTA/COSIC, KULeuven Modeling a MRS signal 1 11 21 31 41 51 61 71 81 91 101 111 121 data point -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 kHz Theoretical model function PR Nijmegen