Research On Glioma Grading Model Based On Multi-sequence Maximum Entropy Discrimination

Glioma is a common brain tumor,and its grade prediction is of great significance to determine the treatment plan and prognosis of patients clinically.Imaging doctors usually classify gliomas based on joint analysis of multiple sequences of magnetic resonance imaging(MRI).The popularity of imaging examinations makes imaging doctors face increasing work pressure.With the maturity and rapid development of computer information technology and big data computing technology,computerassisted diagnosis is widely used in the field of clinical medicine,providing doctors with more efficient professional medical means.At present,there are many research directions for gliomas.In Radiomics research,multi-sequence learning has been used to grade tumors and achieved remarkable results.Multi-sequence learning is a research hotspot of glioma grading.It can analyze glioma data from different angles.Based on the diversity between sequences,accurate diagnosis and grading of multiple gliomas can be achieved.This thesis mainly studies the multi-sequence maximum entropy discriminant model and its improved method to achieve accurate classification of gliomas.In this paper,the public data set Bra TS2017 and Glioma HPPH2018,a data set for glioma in Henan people's Hospital,were used for experimental verification.Both sets of data sets contain four glioma MRI sequences: T1 weighted imaging sequence(T1WI),T2 weighted imaging sequence(T2WI),contrast-enhanced T1 Weighted imaging sequence(CET1)and fluid attenuated inversion recovery(Flair).For two sets of data sets,image segmentation is performed first,then feature engineering is performed on the image data,and finally,multisequence glioma classification model is trained.The main research contents of this article are as follows:(1)In this thesis,a multi-sequence maximum entropy discrimination(MSMED)model is proposed.This model mainly starts from multi sequence,introduces four parameters to construct the maximum entropy discrimination model,and uses the relative entropy as the objective function to solve.The marginal consistency criterion among each sequence and the difference and complementarity of the four sequences are used to improve the grading accuracy of glioma.The experimental results show that the maximum entropy discriminant model based on multiple sequences is better than that based on single sequence.Compared with the average AUC(Area Under Curve)value of T1 WI sequence with the best single sequence training result,the average AUC value of Bra TS2017 and Glioma HPPH2018 data sets on MSMED model is about 13% and 8% higher than that of Med model,respectively.(2)This thesis also proposes a multi-sequence AdaBoost MED(MABMED)model.The MABMED model uses the maximum entropy discriminant model as the weak classifier,and uses AdaBoost algorithm to update four different sequences and sample weights of glioma,and introduces new parameters It is used to represent the identity and value of different sequences.Finally,all the weak classifiers are combined to get a strong classifier to grade gliomas.The experimental results show that under the same conditions,the effect of brain glioma grading of MABMED model is better.The average AUC values of Bra TS2017 and Glioma HPPH2018 data sets in MABMED model training were 0.9485 and 0.9612,respectively,with an average accuracy of 92.57% and 93.19% respectively.