Three-dimensional medical image segmentation method and system based on short-term and long-term memory self-attention model

A three-dimensional medical image segmentation method based on a short-term and long-term memory self-attention model for performing segmentation on a target area image in a medical image, comprising:

• (1) collecting three-dimensional medical images containing target areas and with known target area segmentation golden rules, cropping the original images according to areas of interest, and performing normalization processing on pixel values to obtain training set samples;
• (2) cropping an original image according to an area of interest for an original three-dimensional medical image to be segmented, and performing a normalization operation on pixel values of the cropped image to obtain a sample to be segmented;
• (3) establishing a three-dimensional medical image segmentation network based on short-term and long-term memory self-attention, wherein the network first uses a feature extraction module comprising a plurality of convolution layers, and using a convolution calculation to convert grayscale information in a network input image into feature information to obtain initial image features;
• using a plurality of sequentially connected convolution modules having residual structures to further process the initial image features to obtain corresponding convolution features;
• using a plurality of short-term memory modules based on a multi-head self-attention mechanism to calculate a global correlation between the convolution features of different levels, generating short-term memory features based on the global correlation together with the convolution features used in the correlation calculation;
• performing a global correlation calculation on all of the short-term memory features obtained by a long-term memory module also based on the multi-head self-attention mechanism to generate features to be reconstructed used to reconstruct a segmentation result, wherein the long-term memory module has same structure as any one of the short-term memory modules;
• and finally, performing processing on the features to be reconstructed through a plurality of convolution layers and deconvolution layers by a reconstruction module to obtain a predicted probability of a category to which each pixel of the network input image belongs to realize performing accurate segmentation on the target area in the three-dimensional medical image;
• training the established three-dimensional medical image segmentation network by using the training set samples obtained in (1) and using the corresponding segmentation golden rules as labels, so that the trained network is able to realize performing segmentation on the target area contained in the network input image;
• and
• (4) using the sample to be segmented obtained in (2) as the network input image to input to the trained three-dimensional medical image segmentation network based on short-term and long-term memory self-attention obtained in (3), and outputting a segmentation result of the target area in the sample to be segmented.