Nuclear magnetic image brain gray matter nucleus segmentation method based on convolutional neural network

Abstract

The invention discloses a nuclear magnetic image gray matter nucleus segmentation method based on a convolutional neural network. The method comprises the following steps: preparing and preprocessingdata; constructing global and local branch data pairs; designing a 3D ResUnet network based on the original framework and the residual structure of the 3D Unet network; designing a global and local feature extraction structure, and expanding a network input part into two branches to respectively extract features of global and local data pairs; designing a feature compensation module, and fusing the features of the two branches by using 3D transposed convolution, center clipping and feature splicing operations; designing decoding double branches; inputting the processed data into a final modelfor training; and verifying the segmentation effect of the network on the test set of the pixel-level label, and outputting a segmentation result. The method provided by the invention can greatly reduce the cost of labeling data, provides a data basis for quantitative research of neurodegenerative diseases to a certain extent, and has great significance for exploration of disease pathological mechanisms and clinical treatment.

Description

technical field [0001] The invention relates to a method for segmenting brain gray matter nuclei in MRI images based on a convolutional neural network. Background technique [0002] The morphological changes and iron content of the gray matter nuclei are related to neurodegenerative diseases such as Alzheimer's disease, Parkinson's disease, schizophrenia, multiple sclerosis, for example, the morphology and developmental trajectory of the caudate nucleus, putamen and nucleus accumbens Changes in ASD are associated with Autism Spectrum Disorder (ASD); increases in iron content in gray matter nuclei such as the basal ganglia, globus pallidum, and substantia nigra are associated with Parkinson's disease. At present, the volume and iron content of the inner brain mass are mostly measured by manually drawing the region of interest (Region of Interest, ROI). The measurement of volume and iron is not accurate enough, which affects the diagnosis and evaluation of neurodegenerative d...

AI Technical Summary

Problems solved by technology

Similarly, this method also requires registration of QSM through T1, and requires a wide distribution of training sets, and the expressiveness of Gaussian mixture models is also limited.

In 2017, Jose Dolz et al. designed a 3DFCNN model to segment brain nuclei on T1-weighted images, and also indicated that their method can be used in QSM. The proposed method divides the data into 27 3 image blocks due to 27 3 The size of the image block lacks a lot of global information, and the performance is poor in more complex QSM segmentation tasks

The ICM algorithm can be regarded as a single-layer neural network with limited expressive ability, and the calculation efficiency of this algorithm is low. The original paper pointed out that it takes 1.6±0.5h for an image

Images