Research On Brain MRI Age Estimation Based On Deep Learning

The study of the morphology and function changeof the brainaging process plays an important role in the diagnosis and treatment of neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease.The early ageing of the brain causes the brain age to be different from the actual age of the individual.The estimated age of the brain can provide a preliminary and intuitive representation of the level of brain aging.The MRI imaging technology,especially T1-weighted images,can provide higher-resolution 3D images with better tissue resolving power and is suitable for observing the internal structure of the brain.Therefore,the study of brain age based on MR images is ofgreater importance.The algorithm for estimating the brain age using traditional machine learning methods requires complex preprocessing procedure for brain MR data,such as segmentation and dimension reduction,and depends on the hand-crafted feature,high complexity and large errors are limitations.The existing model of brain ageestimation algorithm based on deep learning is just stacked by several convolution layers.The fitting ability is not strong enough and the estimation error is large.Tosolve these problems,,this paper studies the brain age estimation of MR images based on deep learning framework.This paper first designed a deep learning model for two-dimensional MR slice images to estimate age,namely the ResNet-based brain age estimation model.The residual structure in ResNet can simplify the end-to-end feature mapping into residual mapping and integrate the low-middle-level high-level semantic information well.The input 3D NMR image is divided into multi-channel image slice input network along the Z-axis,feature extraction is performed on the slice scale,middle-low-level high-level semantic information is integrated layer by layer,and finally the age estimation value is output.Simultaneously with the multi-task approach,gender is chosen as the second task to constrain age estimation tasks.In this paper,the algorithm is verified on the open data set.The experimental results show that the proposed algorithm has improved performance compared with existing solutions.Secondly,on the basis of the first work,this paper combines the characteristics of MR data,and can not make good use of the correlation between MR data slices for two-dimensional convolutions,and uses three-dimensional convolution to extract data features to make better use of MR data features in the third dimension.For the shortcomings of training a small sample size,use dense connections and combine feature weighting.At the same time,combined with the domain knowledge of brain age estimation,the high-frequency information of the brain is extracted and sent into the deep network.Experiments in the open library show that the program has a further improvement in performance over the previous scheme.