A Multi-Atlas Image Segmentation Method Based on Orientation and Scale Descriptors

Abstract

The invention relates to a multi-graph image segmentation based on direction and scale descriptors. The method comprises the steps of reading all graph gray images Itrain and corresponding graph mark images L, reading an image Itarget to be segmented, taking the image to be segmented as a reference image and the graph gray images as floating images, registering all graph gray images Itrain to a target image one by one to obtain the deformation fields T (T=(T1, T2 to Tk)) of graph gray image deformation and deformed graph gray images I'train, then allowing the deformation fields to act on corresponding graph mark images, obtaining deformed graph mark images L', using L' prior information, determining a hippocampus dissection position in the image to be segmented, determining an interest area, and calculating the direction and scale descriptors S of all pixels of the graph gray images and target pixels, wherein the pixels in the interest area are the target pixels, and the direction and scale descriptors S are new pixel characteristics. According to the method, the segmentation error caused by uneven brightness between graphs can be reduced, and the accuracy of segmentation is improved.

Description

technical field [0001] The invention relates to the technical field of general image data processing, in particular to a multi-atlas image segmentation method based on direction and scale descriptors. Background technique [0002] Image segmentation methods based on multi-atlas registration combined with label fusion technology mainly include the following segmentation methods: [0003] One: directly use the global similarity of the image as the basis for map selection. The theoretical basis of this method can be understood as: the higher the global similarity, the more similar the grayscale image of the atlas is to the image to be segmented, and the closer the distribution of each tissue structure. This method is simple to operate, but the accuracy of small organ segmentation is low. The global similarity is mainly determined by large organs, and the local similarity of small organs is poorly consistent with the global similarity. Tong found that based on the global simi...

AI Technical Summary

Problems solved by technology

Therefore, selecting maps based on global similarity is not suitable for segmenting small organs

[0004] Two: Using local similarity as a segmentation algorithm for label fusion weights, there are various methods for calculating local similarity. Aljabar et al. use mutual information as a similarity measure, and Tong et al. use the sum of squared residuals as a similarity measure. These algorithms generally They all directly use the local grayscale blocks of the pixels to calculate the local similarity of the pixel features. Although the grayscale image of the map and the grayscale information of the image to be segmented are fully utilized, as well as the spatial structure information of the map, the uneven brightness between the maps is not considered. sex

Different segmentation methods have different ways of constructing dictionaries. Roy uses graphs to train a shared dictionary that can approximate all pixels. Although the dictionary training process is relatively simple, the segmentation accuracy is lower than the segmentation method that trains a unique dictionary for different pixels.

Images