Research On Intelligent Identification For Multiple Sclerosis And Neuromyelitis Optical Spectrum Disorder By Deep Learning Of Magnetic Resonance Imaging

Multiple sclerosis(MS)and neuromyelitis optical spectrum disorder(NMOSD)are both central nervous system(CNS)demyelinating diseases.In the past,NMOSD was considered a subtype of MS.In recent years,with medical progress,MS and NMOSD have been considered two independent diseases in clinics.However,their clinical symptoms and magnetic resonance imaging(MRI)manifestations often overlap,making early differential diagnosis difficult.Drugs such as interferon beta,which treat MS,can aggravate the condition of NMOSD patients.Therefore,early identifying the two diseases and quantitative monitoring of the lesions are of great clinical significance.Magnetic resonance imaging is a standard method for CNS diagnosis and plays a vital role in clinical diagnosis and research.Brain MRI can objectively reflect the morphology and distribution characteristics of multiple lesions in the brain and plays an essential role in disease diagnosis,disease course monitoring,and prognosis evaluation of MS and NMOSD.MS and NMOSD are most likely to affect the myelin sheath of the white matter in the brain and then show high signal lesions in T2 weighted imaging(T2WI)or T2 fluidattenuated inversion recovery(T2-FLAIR),that is,white matter hyperintensities(WMH).Three-dimensional(3D)MRI examination sequence has been gradually applied to clinical practice in recent years.It has dramatically improved the detection rate of lesions,making it possible to make the quantitative diagnosis of MS and NMOSD brain WMH based on artificial intelligence(AI)and Artificial intelligence differential diagnosis based on automatic segmentation of WMH.This doctoral dissertation focuses on the intelligent auxiliary diagnosis of multiple sclerosis and neuromyelitis optical spectrum disorder based on 3D brain MRI images.In this dissertation,the automatic segmentation methods of white matter high signal region,T1 weighted imaging(T1WI)enhanced focus signal region,and two intelligent aided classification and diagnosis methods of diseases in 3D brain MRI images are studied.This dissertation carries out research work from the following three aspects:1.It is necessary to obtain the characteristics of white matter hyperintense lesions in MS and NMOSD brain MRI images.Therefore,this paper proposes a 2.5D deep learning network model(CBAM-based Dense UNet,CD-UNet)based on multimodal brain MRI images to extract and segment the WMH in MS and NMOSD brain MRI images.2.5D CD-UNet is based on the Fully Convolutional Dense Net,(FC-Dense Net),and embeds the convolutional block attention module(CBAM)in the attention mechanism.2.5D CD-UNet is more suitable for automatic segmentation of small-scale data such as multimodal brain MRI images of MS and NMOSD.This also reduces the hardware dependency of the 3D deep learning model to a certain extent.The experimental results show that the 2.5D CD-UNet proposed in this dissertation can accurately and efficiently segment the white matter high signal in MS and NMOSD brain MRI images,and the effect is obviously better than many classical deep learning network models.2.Compared with traditional MS and NMOSD brain MRI images,the amount of data of T1WI-enhanced lesion images is less,and the white matter of lesions is more sparse and scattered,which is more difficult to distinguish.Therefore,higher requirements are put forward for intelligent automatic identification of related images.Given the above problems,based on optimizing FC-Dense Net,the 2.5D deep learning network model is first trained and learned on T2-FLAIR traditional MS and NMOSD brain MRI data.Secondly,the trained model was used as a pre-training model for the automatic segmentation of brain WMH of T1WI-enhanced lesions.Finally,its parameters were used to train T1WI-enhanced lesions to establish a transfer learning model.The experimental results show that the migration model based on the 2.5D deep learning network proposed in this dissertation has superior performance in the T1WI-enhanced WMH segmentation task,which is significantly higher than many classical deep learning network models.To a certain extent,it solves the problems of the small amount of T1 WI enhanced focus data and the difficulty of segmentation of hyperintensities in related white matter.3.Multiple sclerosis and neuromyelitis optical spectrum disorder overlap in the white matter hyperintensity areas of brain MRI images,so it is difficult to identify the two images accurately.In this paper,the classification of MS and NMOSD brain MRI images was studied after the automatic segmentation of the white matter hyperintensity data was processed.Aiming at the problems of large data volume,many parameters,and high resource consumption in 3D deep learning models,this paper proposes a 3D data compression module strategy to map 3D brain MRI images from high-dimensional to low-dimensional 2D single slice data.Then,the2 D slice data is input into a 2D deep learning network model,which finally outputs the classification results of MS and NMOSD.In the case of limited data,the 2D deep learning network model based on Image Net’s pre-trained model further improves the effect.The experimental results show that the effect of the method proposed in this dissertation is significantly better than that of other classical deep learning methods.The parameter quantity and effect are also significantly higher than the model directly using 3D deep learning.