Prediction Of MOGAD In Acute Disseminated Encephalomyelitis Based On Multimodal Magnetic Resonance Features

Idiopathic inflammatory demyelinating diseases(IIDDs)are a group of immunerelated central nervous system demyelinating diseases characterized by demyelination and inflammatory cell infiltration.Among the many subtypes of this disease spectrum,Acute disseminated encephalomyelitis(ADEM)has the characteristics of acute onset,long duration,and susceptibility in children.Another subtype of demyelinating disease with similar clinical symptoms and imaging findings to this disease is Myelin oligodendrocyte glycoprotein(MOG)antibody-related disease,which is a young but widely recognized disease and has become a research hotspot in the field of demyelination in recent years.Prior to the independent definition of MOGAD,multiple studies showed that MOG antibodies had a significant impact on the course of ADEM patients,and MOG antibody-positive ADEM patients were more likely to have multiple courses.Nowadays,the differentiation between typical ADEM and MOGAD with ADEMlike manifestations mainly relies on the detection results of MOG antibodies.However,this method is not always reliable and efficient.In reality,only a few companies and research institutions have the capability to test for MOG antibodies,and most tertiary medical centers cannot send patients’ blood or cerebrospinal fluid samples to testing facilities in a timely manner,which poses potential risks for ADEM patients with acute onset characteristics.A MOGAD prediction model with immediacy,easy operation,and quantifiability could effectively solve the above problems and guide the dosage of steroids in the acute phase of patients.To address the above issues,this thesis introduces machine learning methods for rapid and effective identification of MOGAD in pediatric patients with ADEM-like symptoms.Specifically,the following tasks are included:(1)This thesis proposes a machine learning-based MOGAD identification algorithm using radiomics features extracted from FLAIR sequences.The method based on feature engineering can ensure the accuracy of the model while providing interpretability.First,the study cohort was divided into four subgroups based on age(6years old as a cut-off)and gender,and machine learning classification models were separately built for each subgroup.Radiomics features(1039 dimensions)were extracted for each lesion(2D connected domain),and dimensionality reduction was performed using redundancy analysis,significance testing,and LASSO screening.Then,the selected features were grouped and their frequency was calculated to provide interpretability for the model.Four machine learning models were built using the selected features and their performance was validated with AUC ranging from 89.0%to 99.2% across the four subgroups.Additionally,to improve the predictive performance,the same process was applied to T1 and T2 sequences,resulting in improved performance for the 0-6-year-old subgroups.(2)This thesis constructed a multi-modal deep segmentation network for automatic lesion segmentation.Specifically,to address the problem of imbalanced lesion information across MR sequences in the multi-modal model,we proposed a method of using a BILSTM network to learn the grayscale range of the lesion,generating a weight matrix that contains most of the FLAIR sequence lesion information,and using this weight matrix to enhance the other two non-significant modalities.Then,based on a multi-modal U-Net network,we performed lesion segmentation using the enhanced images.The experimental results showed that compared to using a single-modal U-Net and a non-enhanced multi-modal model,our method has better performance.(3)A MOGAD prediction model was constructed based on the two models mentioned above,and a pediatric ADEM-like patient MOGAD identification system was developed using Matlab GUIDE tool based on this model.The system can provide data preview,gray level histogram prediction result display,lesion segmentation,and MOGAD prediction functions.This system can assist clinicians in identifying MOGAD patients in patients with ADEM-like symptoms during acute onset before sample testing.