This work could benefit neurosurgeon and neuroradiologists in need of a support decision system.The non-invasive classification and grading of brain tumor would benefit patients a lot, especially those in need of pathological section to guide the following treatments. There have been many researches on the classification or grading of brain tumor, but barely no work in a comprehensive system combine classification and grading together. Therefore, we designed an automatic multi-layer classification system to simulate the brain tumor diagnosis process according to clinic workflow: 1).To indentify a brain tumor based on quantitative MR spectroscopic (MRS) parameters; 2).To diagnose the brain tumor type based on T1W/T2W MRI, MRS and clinical information; 3).To indicate the grade of brain tumor based on T1W/T2W MRI and MRS. Support Vector Machine (SVM)¹ was used in the Step 1 and 3, and a Tumor Model was established in Step 2. The whole three steps together constituted an automatic multi-layer classification system. Subjects: Please refer to Table 1 for clinical data used in this research. Data-driven SVM was used in Step 1 and 3,, whereas an algorithm-driven Tumor Model was established in Step 2. Without the need of training data, data for Step 1 and 3 can all be used in Step 2. Sequence: Each subject underwent T1W, T2W and 2D MRS (TR=2000ms, TE=144ms, voxel=10×10mm²) scan with volumes of interest (VOI) covering lesion, edema and healthy areas. All scan was underwent respectively at a 3T MRI scanner (Achieva 3.0T TX, Philips Medical Systems, the Netherlands). Feature extraction: Pre-processing and quantification of MRS data was running on LCModel (a software professional in MR spectra processing) before feature extracting. The input features for each step are given in Table 2. We designed a MATLAB Toolbox (Fig. 1) to extract the quantification features of different areas. The color is displayed according to the concentration of Cho/Cr, where the red regions represent suspected tumor lesion with high Cho/Cr. The three regions are manually selected by the color information and with the help of experienced neuroradiologists. Classification and grading: 33 and 18 subjects for Step 1 and 3 were used respectively for SVM, including machine learning, parameter optimization and dimension reduction. Because of the lack of data, we focused on glioma grading in Step 3.In Step 2’s Tumor Model, firstly, we collected a set for each tumor (G for glioma, M for metastasis, O for meningioma and others) according to the clinical information, as well as a set, P1, for each subject. Then, compared P1 with G, M or O to calculate whether P1 was a subset of them. If it was, we can have the first score for each tumor: Score_G1=length(P1)/length(G); Score_M1=length(P1)/length(M); Score_O1=length(P1)/length(O). Secondly, we collected a MRS set for each tumor (MRS_G for glioma, MRS_M for metastasis, MRS_O for meningioma and others) with mean value² of the NAA/Cr, Cho/Cr and Ins/Cr, and a MRS set, P2, for each subject. Pearson Correlation Coefficient(R) was calculated between P2 and each tumor set to get the second score: Score_G2=R(MRS_G,P2); Score_M2=R(MRS_M,P2); Score_O2=R(MRS_O,P2). Finally, each tumor had a final score: Score_G=Score_G1+Score_G2; Score_M=Score_M1+Score_M2; Score_O=Score_O1+Score_O2. Tumor with the highest score was considered as the diagnosed type. And thus, a multi-layer classification system is established, as shown in Fig. 2.The classification performance of each step and the whole system are summarized in Table 3. The system can clearly distinguish tumor and non-tumor patients (e.g. inflammation, multiple sclerosis and etc.), different tumor types and grades with an overall accuracy of 90%. The accuracy of each step is also quite acceptable. Its comprehensive performance can be useful in clinical work. we designed a comprehensive multi-layer system for clinic use. Step 1 has an accuracy of 100%, however, it has 28 support vectors (SVs). It may be caused by the similarities between tumor and non-tumor, or over-fitting during machine learning. Step 3 has an accuracy of 100% with 12 features and 13 SVs. Principle Components Analysis may reduce the number of features and SVs to find the optimal features. The mean value used in Tumor Model should be updated with increasing dataset. A more complicated model could improve the accuracy and help to diagnose more tumors. In our future work, different classification algorithms and data quality control would help improve the overall system performance. This study proposed a novel multi-layer classification and grading system of brain tumor. It takes in many features that were easily ignored, and has reached an overall accuracy of 90%. It could be a helpful auxiliary tool for neurosurgeon and neuroradiologists by reducing diagnosis uncertainty.