Research And Application On Brain MRI Image Analysis Algorithm For Migraine Without Aura

Objective:Migraine without aura(MwoA)is one of the common primary headache disorders in the clinic,and its clinical diagnosis is mainly based on the international classification of headache disorders(3rd Edition)criteria combined with the clinical manifestations of patients.Currently,magnetic resonance imaging(MRI),as a novel MwoA imaging diagnostic tool that can provide brain functional imaging,plays an increasingly important role in clinical research and auxiliary diagnosis of MwoA.However,subject to physician experience and image analysis methods,the diagnostic accuracy of existing MRI-based methods for MwoA needs to be improved.Therefore,research on intelligent image analysis methods to assist clinical diagnosis of MwoA,has become an urgent problem in the field of medical artificial intelligence.Machine learning based functional connectivity analysis methods on brain MRI images,is one of the mainstream directions in the current research of MwoA intelligence assisted diagnostic algorithms.However,there are some problems in this type of methods,such as the reliance on predefined brain atlas templates when extracting functional connectivity features,the overall level of intelligence and low accuracy of the method.In this paper,the MwoA brain MRI image analysis algorithm is deeply studied,and two practical algorithm models are proposed to advance the clinical auxiliary diagnosis of MwoA.Methods and results:For the problems of strong subjectivity in functional connectivity extraction,low overall intelligence and accuracy of the algorithm,the MwoA3D-Net algorithm framework based on the 3D convolutional neural network is proposed.Group information guided independent component analysis is introduced into MwoA aided diagnosis task,which is used to generate eight resting-state brain networks of subjects to avoid the difference of results caused by different prior templates.The RSN-Net module can automatically extract the 3D spatial structure features of restingbrain network and output diagnostic information.After fusing the diagnostic information,the designed full connection module is used to complete the auxiliary diagnostic task.In addition,optimization strategies such as data enhancement,LI regularization,and L2 regularization are incorporated in the algorithm,which can effectively suppress the overfitting problems.Experimental results on a dataset of 60 MwoA and 65 healthy subjects showed that the average accuracy of the MwoA3D-Net algorithm was 98.40%.Aiming at the problem that the abnormal situation of eight resting-state brain networks is different,a weakly supervised learning algorithm framework with 3D deep multiple instance learning is proposed,dubbed R3D-DMILSAM.The innovation of the method is that the eight resting-brain networks generated by each subject are encapsulated into a bag,and the patient information is used as the label of the bag.Then the MwoA aided diagnosis problem is transformed into the bag classification problem in multiple instance learning.The deep instance generation module designed by algorithm can extract the 3D spatial structure features of resting-brain network and automatically generate a series of deep instance.In addition,to explore the resting-state brain network abnormalities of patients,a spatial attention pooling layer is incorporated into the algorithm,which can automatically weigh the resting-state brain network abnormalities.The addition of this pooling layer further improves the diagnostic accuracy of MwoA.According to Bernoulli distribution,the bag semantic representation can be transformed into the bag category information to complete the MwoA aided diagnosis task.Experimental results show that the accuracy of the algorithm can reach 88.80%.Conclusion:In the aspect of intelligent image analysis aided MwoA diagnosis,the MwoA3D-Net and R3D-DMILSAM algorithm frameworks proposed in this paper have achieved high diagnostic accuracy and good research and application value.Studies have found that resting-brain networks such as the left frontoparietal network can serve as MwoA potential biomarkers for individualized diagnosis,and they provide important evidence for studying the pathological mechanisms of MwoA.This paper not only provides an important technical method for brain MRI image analysis and disease research,but also has some borrowing and pushing significance for the auxiliary diagnosis of brain diseases.