Research On Computer-Aided Diagnosis Models For Brain Diseases Based On Magnetic Resonance Imaging Using Deep Learning Techniques

The brain,as a sophisticated organ within the human body,is characterized by complex structure and diverse functionality,serving as the origin of human thought and controlling human consciousness.Brain diseases have a high incidence,disability rate,and mortality rate,however,the current prediction and identification of brain diseases mainly rely on clinical symptoms,which require a large number of experienced doctors,but the current medical resources are still insufficient.Magnetic Resonance Imaging(MRI)technology has been widely used in clinical practice and provides data support for understanding the brain’s working mechanism and conducting clinical auxiliary diagnosis.With the development of artificial intelligence,analyzing MRI data based on deep learning technology and developing accurate Computer-Aided Diagnosis(CAD)models have important application value.Therefore,this study investigates deep learning CAD models based on MRI and proposes novel methods for subtype diagnosis of patients with Mild Cognitive Impairment(MCI),aimed at achieving a three-year conversion prediction from MCI to Alzheimer’s Disease(AD)and assisting in the diagnosis of Major Depressive Disorder(MDD).Currently,the following problems still exist in this field of research: due to the limitations of MRI acquisition cost,the sample size of MRI available in most of the studies is small,but the large number of raw MRI features easily leads to overfitting and affects the model performance,and the interpretability of the models is often overlooked,making it difficult to understand the outputs;the existing methods suffer from limitations in effectively fusing multimodal MRI data,which in turn affects the accuracy and reliability of the results.Therefore,it is important to propose new models and research frameworks to address these issues.The research work in this thesis consists of the following two parts:Research work 1: To alleviate the issue of overfitting caused by an excessive number of original MRI features but few available neural imaging samples,we extracted various cortical morphological features that contain regional brain-level information while having a lower number of features compared to the original images.We then propose a new transformer-based model to fuse these features for predicting the three-year conversion from MCI to AD.The model achieved an accuracy of 83.3%on the test dataset,indicating its clinical application value.Additionally,we conducted occlusion analysis on the input features,masking them based on brain regions and feature types,to assess their contribution to MCI conversion prediction.The results showed that masking the caudal anterior cingulate and pars orbitalis based on brain regions had the greatest impact on the model’s accuracy,suggesting that these brain regions make the greatest contribution to our model’s prediction of MCI conversion and should receive more attention in clinical applications.Moreover,masking cortical volume and thickness based on feature types revealed that they may be more reliable cortical morphological features for predicting MCI conversion.Research Work 2: In this thesis,we propose an adaptive multimodal MRI fusion model to fuse multimodal MRIs more effectively,and the model is validated in an MDD-assisted diagnostic task.The model first designs the Brain Function Encoder and the Brain Structure Encoder to extract functional and structural information of the subjects’ brains from multi-modal MRIs,respectively.Then,it uses a co-attentive fusion classification network to capture the interaction between functional and structural information and adaptively learns intermediate representation features for classification.The model was validated to achieve an accuracy of 75.2%.This study also analyzed the attention values in the co-attentive fusion network and found that the model paid significantly more attention to structural features in the MDD group than in the healthy controls group,suggesting that there may be more information in brain structure than brain function that is highly relevant to MDD.Through the above research,this thesis has explored the development of deep learning CAD models for brain diseases based on MRI,and proposed solutions to some of the existing problems in this field.The model developed also holds reference value for predicting and identifying other brain diseases.