The Design Of Brain Tumor Classification Based On Multi-Model

Recently MRI and MRS have been an effective tool for aiding the radiological diagnosis of brain tumor. MRI can provide three dimensional anatomic image with high spatial resolution, and about 30-90% accuracy can be achieved based on MRI. It varies by the type and grade of brain tumor. The MRS shows the metabolite information of lesions. It has been reported that there are obvious difference of spectra between tumor and normal tissue.In this study, our purpose is to evaluate whether we can get a better predictive accuracy by combining long TE MRS data and PWI feature using LDA and SVM. We parsed DICOM image and obtained registration of MRS multi-voxel and MR image. The multi-voxel spectra of VOI was quantized by LCModel. We consider four classification groups:normal tissue, low-grade glioma, high-grade glioma and metastasis. LDA and SVM were performed on the selected metabolite concentration of the original spectra and rCBV of PWI. The input parameters were selected based on the characteristics of each brain tissues. In general, this study obtains high performance for classifying the four categories, the classifiers for the normal versus tumors and metastasis versus gliomas achieved above 90% accuracy, even 100%. For the low-grade versus high-grade glioma, the classification accuracy with rCBV and metabolites concentration as input features increases 10 percent compared to using metabolites only, and achieved obove 80% accuracy.Finally, we designed the database for the brain tumor decision support system. The classifiers were applied into this this system and we designed a user friendly interface based on Java. This software is completely based on research and practical application development.