Contour perception multi-organ segmentation network construction method based on class-by-class convolution operation

Abstract

The invention discloses a contour perception multi-organ segmentation network construction method based on class-by-class convolution operation. The contour perception multi-organ segmentation networkconstruction method comprises the following steps: 1, performing region coarse segmentation and edge detection on multiple organs of the abdomen; 2, introducing a semantic-guided class-by-class multi-scale attention mechanism; step 3, performing class-by-class fusion of multi-branch information; step 4, performing introduction of multi-task loss. According to the invention, the advantages of a convolutional neural network and a gated recurrent neural unit are utilized, and for the characteristics and difficulties of a multi-organ segmentation task, via the contour information assisted multi-scale feature extraction, a class-by-class multi-scale semantic attention mechanism, a class-by-class cavity convolution fusion mechanism and a plurality of loss functions can be introduced to relievethe inter-class imbalance problem of organs; multi-organ segmentation is performed on a three-dimensional CT image more efficiently and accurately, and the advantages of the invention are verified ona data set containing 14 types of organ labels; the invention can be widely applied to computer-aided diagnosis and treatment application, such as endoscopic surgery, interventional therapy and radiotherapy plan making.

Description

technical field [0001] The present invention relates to a method for constructing a contour-aware multi-organ segmentation network based on a class-by-class convolution operation, especially a contour-aware three-dimensional convolutional neural network based on a class-by-class multi-scale semantic attention mechanism and atrous convolution fusion, which is used for The invention relates to multi-organ segmentation of abdominal CT images, belonging to the technical fields of digital image processing, pattern recognition and medical imaging engineering. It mainly involves convolutional neural network (CNN) and gated recurrent neural unit (GRU), which can be widely used in computer-aided diagnosis and therapeutic applications, such as endoscopic surgery, interventional therapy, and radiotherapy planning. Background technique [0002] In the existing medical diagnosis process, manual methods still cannot meet patients' requirements for accuracy and timeliness. The results of ...

AI Technical Summary

Problems solved by technology

In addition, the existing segmentation methods are only simple imitation and extension of 2D image segmentation methods and 3D single-organ segmentation methods, without structural optimization for the characteristics and difficulties of 3D multi-organ segmentation: First, in the multi-organ segmentation task, the target The number of organs generally does not exceed fifteen. Compared with natural image segmentation data sets, the complexity is greatly reduced; secondly, in the segmentation task of natural images, a target may appear in any position or multiple positions of the image, and in each In a group of 3D CT images, the number and location of organs are highly predictable; in addition, medical image segmentation has higher requirements on the effect compared with natural images, especially at the edge of organs

Images