Three-dimensional U-Net brain tumor segmentation method fusing conditional randomness and residual errors

Abstract

The invention provides a three-dimensional U-Net brain tumor segmentation method fusing conditional randomness and residual errors. The three-dimensional U-Net brain tumor segmentation method comprises the steps: carrying out the three-layer cascade network architecture training of a training set, and obtaining a model of a convolutional neural network; testing the test set through a convolutionalneural network model to obtain a probability matrix corresponding to each category of the brain tumor; and carrying out post-processing on the probability matrix, and updating the probability to obtain a final brain tumor segmentation result. The three-dimensional U-Net brain tumor segmentation method simplifies the problem step by step, and obtains a better effect. compared with a traditional method, the three-dimensional U-Net brain tumor segmentation method has more advantages that the adopted network has better learning ability, and a residual block is introduced to reduce the influence of gradient explosion, gradient disappearance and network performance degradation caused by deepened network layers; and information of a three-dimensional space is used to the greatest extent.

Description

technical field [0001] The invention relates to the technical field of digital image processing, and relates to a processing method of brain magnetic resonance images, and more specifically, relates to a three-dimensional U-Net brain tumor segmentation method combining random conditions and residual errors. Background technique [0002] The tumor segmentation of brain magnetic resonance images is an international competition called BraTS (Brain Tumor Segmentation) every year. This competition provides clinical data sets to participants every year. For example, in the public BraTS2015 dataset, all pictures in the dataset are magnetic resonance images (MRI). Prior to this, the analysis and processing of these huge and cumbersome magnetic resonance image (MRI) data sets was done manually by doctors or professional researchers to extract the location of the tumor and the composition of the tumor, so that manual analysis will not only consume A lot of manpower and material resou...

AI Technical Summary

Problems solved by technology

However, these methods also have defects of one kind or another. For example, although the deepening of the network layer can bring better segmentation results, the problems of gradient explosion, gradient disappearance and network performance degradation cannot be ignored.

Images