Research On Magnetic Resonance Imaging Based On Deep Learning

The human brain is an extremely complex structure composed of hundreds of billions of different types of neurons.Trying to understand the mechanism of the human brain is the ultimate challenge for humans to pursue natural laws and self-awareness.Brain science is dedicated to studying and analyzing the structure and function of the nervous system,revealing the laws of various neural activities,elucidating its mechanisms at various levels,and preventing,diagnosing and treating neurological and mental disorders.MRI technology has been widely used in brain science research because of its non-invasive advantages.Most of the scientific research results today are based on MRI data.Based on the deep learning algorithm,this paper has conducted in-depth research on two important topics in brain science which is brain tumor segmentation and visual information decoding.Among them,brain tumor segmentation is very important in the diagnosis of brain diseases.Before formulating appropriate and effective treatment plans for patients with brain tumors,doctors need to use MRI to determine the location and size of the tumor in the brain.However,it is time-consuming and laborious to directly divide the specific location of the tumor manually,and ordinary people cannot do this job.Therefore,the development of a concise and efficient automatic brain tumor segmentation algorithm is of great help to clinical medicine.On the other hand,visual neuroscience is an extremely important branch in the field of cognitive neuroscience.As the eyes are the windows of the soul,scientists have done a lot of research on the visual information processing mechanism of the brain.Decoding visual information helps us understand the mechanism of the brain’s visual cortex and provides a biological theoretical basis for the field of computer vision.In these two fields,this paper proposes a brain tumor segmentation algorithm based on VAE skip connection and a visual information decoding algorithm based on improved convolutional features.The main work and the corresponding conclusions are summarized as follows:1.Based on the 3D U-Net brain tumor segmentation model combined with the variational autoencoder,a structure called VAE jump connection is proposed,which cleverly uses the autoencoder and U-Net in the model.The similarity on the above combines the position information of the VAE branch decoder and the semantic information of the U-Net decoder to improve the segmentation accuracy of the network.In addition,due to the high cost of labeling medical images and fewer datasets,models based on deep learning often have overfitting problems.In order to solve this problem,this paper proposes to use the Shake Drop module embedded in the residual block to regularize the network.Experimental results prove that this method can reduce the overfitting to some extent and improve the segmentation performance of the network.2.Based on an algorithm called Generic object decoding,pointed out that the convolutional features extracted by the algorithm lack geometric invariance,and use those feature with stronger geometric invariance,like the spatial max-pooling feature and cross-dimensional weighting feature,to improve it.The experimental results show that the proposed two improved convolution features obviously help to improve the recognition performance of Generic object decoding algorithm.At the same time,compared with the original randomly extracted features,the spatial max-pooling feature and the cross-dimensional weighting feature have better robustness.In addition,this paper proposes to take logarithm to make the feature distribution close to the Gaussian distribution,which is of great help to the training of the decoder contained in the algorithm.