Deep Learning From Diffusion Tensor Image For Parkinson's Disease Detection

Parkinson's disease is a brain disease that causes shaking,stiffness,and difficulty walking,balancing,and coordinating.The symptoms of Parkinson's disease patients usu-ally start gradually in old age and worsen over time.As the disease progresses,people may have difficulty walking and talking.They may also have mental and behavioral changes,sleep problems,depression,memory impairment and fatigue.This disease has brought a heavy economic burden to the people and the country.Nowadays,machine learning,especially deep learning technology,has been widely used in fields such as autonomous driving and face detection.Research shows that related technologies can discover the differences between tiny features that the human eye cannot obtain,and in many ways The accuracy of medical image classification tasks has exceeded that of high-level human experts.The current image diagnosis technology based on ma-chine learning is mainly based on magnetic resonance imaging(MRI),and its results are difficult to interpret and have low robustness,and cannot be applied to practical tasks.Therefore,this study uses diffusion tensor imaging(DTI)that can reflect functional data of the brain,and proposes a new computer-aided diagnosis framework for subregional integrated Parkinson's disease based on convolutional neural networks.This framework first proposes to cut the brain diffusion tensor image data into a standard 116 brain network map area(that is,116 brain regions),and then use deep convolutional networks to super-vise and train 90 brain regions except the cerebellum to obtain 90 Deep learning models.The next step is to use the greedy algorithm to screen the sub-region model,and finally the weighted average of the results of the selected region combination is the final result.This research has the following innovations: 1)use diffusion tensor images to reflect brain functional activity data for diagnosis? 2)the framework can provide more intermediate in-formation than the previous end-to-end method,that is,which areas display the Patients are at risk of illness.3)Current research shows that the robustness of the model is ap-proximately inversely proportional to the classification accuracy.The single sub-region model with low accuracy is more robust than the current end-to-end model,and then it is integrated to compare with the end-to-end model to have better classification effect and stronger practical application.value.4)The results show that the framework has achieved88 % of the cross-validation set accuracy rate for the diagnosis of Parkinson 's disease,and the recognition accuracy rate in the randomly selected 100 sets of test sets without training is 83%.This result is far superior to the diagnosis of human experts.In addition,experiments show that the performance of this sub-regional integration framework is su-perior to the diagnosis method using end-to-end network.5)Through visualization of the results,this study points out the key areas and key points of network diagnosis,and the results may provide new information for the medical community to understand the dis-ease.6)A new counter-attack mechanism for medical tasks is proposed,which can help the medical model to verify and improve its robustness under simulated actual conditions.