Research On Deep Learning-Based MRI-Assisted Diagnosis Of Hepatolenticular Degeneration

Hepatolenticular degeneration(HLD),also known as Wilson disease(WD),is an extremely rare autosomal recessive monogenic disease caused by decreased copper excretion in the body.It is one of the few neurogenetic diseases that can be diagnosed and treated,and the key is early detection,diagnosis and treatment.However,the variability of the first symptoms leads to the fragmentation of consultation departments,such as neurology,gastroenterology,and surgery,resulting in misdiagnosis and omission.Coupled with the fact that early lesions are mild,clinical symptoms are not obvious,and medical personnel are not sufficiently aware of the disease,it is difficult for hospital outpatient clinics to achieve early and accurate identification and diagnosis of preclinical patients.Magnetic Resonance Imaging(MRI)neuroimaging technology can provide different information on brain structure and function,and for HLD,MRI is the best means of HLD diagnosis and even differential diagnosis,but because of the rarity of HLD,radiologists in many local hospitals are not equipped to diagnose this disease.Classification of medical images using deep learning techniques is one of its important applications in the medical field.If deep learning technology can be used to assist physicians in the early stages of HLD onset to accurately diagnose and effectively treat MRI medical images of HLD patients,HLD patients can often achieve the same life and lifespan as healthy people.This is an extremely important research value and significance for early screening of HLD.This thesis focuses on the classification study of HLD samples and non-HLD samples of the constructed 3D MRI medical image dataset,and builds a deep learning based binary classification model,the main research work is as follows:(1)The HLD dataset was constructed from 3D MRI medical image samples provided by hospitals.The image samples were obtained from the case information of patients with HLD and non-HLD diagnosed and collected from March 2011 to October 2021.After a series of data processing,a positive sample size of 1120 was constructed,which is the largest sample size dataset related to HLD and can provide a basis for subsequent studies on HLD.(2)A migration learning-based approach is proposed to construct a binary classification network model and conduct a study for the HLD dataset.Age bias was found in the acquisition of non-HLD samples and HLD samples.In order to eliminate the influence of age characteristics of HLD medical imaging samples as much as possible,age sampling was performed three times,and finally a subdataset of HLD with a total sample size of 2040 was reconstructed.(3)To address the small number of samples in the reconstructed HLD sub-dataset,the low classification accuracy of samples without obvious abnormalities in HLD,and to enhance the learning ability of residual network features,the thesis successively proposed the application of data augmentation and a multi-head self-attention mechanism-based approach.The performance of the binary classification model was further optimized for the sub-dataset of HLD,and the classification accuracy of samples without significant abnormalities in HLD was finally improved from 33.33% to66.67%,which is helpful for the early screening of HLD.