CN113870243A

CN113870243A - Magnetic resonance brain image hippocampus segmentation method

Info

Publication number: CN113870243A
Application number: CN202111189822.5A
Authority: CN
Inventors: 郑强; 刘彬; 童向荣; 武栓虎
Original assignee: Yantai University
Current assignee: Yantai University
Priority date: 2021-10-11
Filing date: 2021-10-11
Publication date: 2021-12-31

Abstract

A magnetic resonance brain image hippocampal segmentation method comprises the following steps: linearly registering a nuclear magnetic resonance brain image to be segmented on a template; step 2: cutting each nuclear magnetic resonance brain image into image blocks containing the hippocampus by a determined box which is enough to cover the hippocampus in the space; and step 3: filling image blocks by using zero filling, inputting the image blocks into a designed deep learning network model HPCSeg-Net, outputting a probability matrix of the sea horse through a softmax function in the network model, and 4: and (4) inputting the hippocampal probability matrix obtained in the step (3) into an argmax function to convert the probability matrix into a 0-1 binary label, and then reversely filling the 0-1 binary label to obtain a hippocampal segmentation result. The invention has the characteristics of high segmentation speed and high segmentation precision.

Description

Magnetic resonance brain image hippocampus segmentation method

Technical Field

The invention belongs to the technical field of brain image hippocampus segmentation, and particularly relates to a magnetic resonance brain image hippocampus segmentation method.

Background

Hippocampal lesions are the causative factors of various mental diseases, and hippocampal segmentation is an important prerequisite for further analysis of diseases by using images, and its volume, shape characteristics, and the like have also been widely used in analysis of mental diseases such as alzheimer's disease. However, the magnetic resonance image has small bilateral hippocampus volume, low contrast with peripheral tissue structures, and unclear boundaries. Therefore, accurate and fast segmentation of hippocampus from MRI remains a challenge.

In the prior art, a multi-atlas method is more stable in segmentation of structural changes, but the segmentation algorithm excessively depends on the performances of an image registration algorithm and a fusion algorithm, accurate segmentation of the hippocampus with large differences is difficult to realize by means of a simple registration algorithm, and the segmentation time is increased along with the increase of the registration number.

A semantic gap problem can be generated based on a traditional U-Net model, and the class imbalance in the sea horse segmentation also enables a deep learning network not to extract enough fine-grained information.

Disclosure of Invention

In order to overcome the technical problems, the invention aims to provide a magnetic resonance brain image hippocampal segmentation method which can extract fine-grained information and weaken the semantic gap problem at the same time.

In order to achieve the purpose, the invention adopts the technical scheme that:

a magnetic resonance brain image hippocampal segmentation method comprises the following steps;

step 1: in order to reduce the differences in the individualization space, the MRI brain image to be segmented is linearly registered to a voxel size of 1 x 1mm³On the MNI152 template;

step 2: determining a box which is enough to cover the hippocampus in the space, fusing the box by using the labels of the selected atlas images, determining the boundaries of the box, expanding 7 voxels outwards in each direction, and cutting each nuclear magnetic resonance brain image into image blocks which are 60 x 48 in size and contain the hippocampus;

and step 3: filling the image blocks with the size of 60 × 48 into 64 × 48 by using zero filling, inputting the image blocks into a designed deep learning network model HPCSeg-Net, and outputting a probability matrix of the hippocampus through a softmax function in the network model, wherein the formula is as follows:

wherein z is_iThe output value of the ith node is C, and the number of the classified classes is C;

and 4, step 4: and (4) inputting the hippocampal probability matrix obtained in the step (3) into an argmax function to convert the probability matrix into 0-1 binary labels, and then back-filling the 0-1 binary labels to obtain a hippocampal segmentation result with the size of 60 x 48.

The invention has the beneficial effects.

The invention designs a cascade automatic focusing attention mechanism comprising an automatic focusing module and an attention module based on a basic framework of U-Net, and uses the cascade automatic focusing attention mechanism in U-Net jump connection to reduce semantic gap problems.

The standard convolution in U-Net is replaced by an adaptive feature reorganization and recalibration module to extract fine-grained information.

The HPCSeg-Net provided by the invention is superior to the MAIS method in the aspects of segmentation performance and calculation cost:

the HPCSeg-Net segmentation of 70 magnetic resonance images only needs 0.35 minute, the average segmentation time is less than 1 second, and the segmentation result is superior to the MAIS method. While the average run time of the MAIS method segmentation included NLP (3.73 minutes), RLBP (10.95 minutes), ML (46.18 minutes), LLL (4.8 minutes), RF (46.7 minutes), and RF-SSLP (48 minutes).

Description of the drawings:

fig. 1 is a schematic diagram of a magnetic resonance brain image hippocampal segmentation method of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1: the invention provides a hippocampus segmentation network model HPCSeg-Net based on a U-shaped convolution neural network framework, which adopts a cascade automatic focusing attention mechanism and a self-adaptive characteristic recombination and recalibration module. A cascading autofocus attention mechanism includes an autofocus module and an attention module for a jump connection between down-sampling and up-sampling to reduce a semantic gap. The autofocus module generates stronger features (number of convolutional layers 3, 2, 1 in autofocus modules 1, 2, 3) by parallelizing multiple convolutional layers of different expansion ratios (2, 4, 6), and the follow-up attention module contains 2 inputs, where the up-sampled feature map contains deeper level information, which can suppress irrelevant areas around the hippocampus, highlighting salient features. The feature recombination adopts linear expansion and compression to generate complex semantic segmentation features, and the feature recalibration module collects context information and retains spatial information. In order to prevent the network from being over-fitted, a structured packet loss form is added in the feature recombination and recalibration module to standardize the convolution network.

Claims

1. A magnetic resonance brain image hippocampal segmentation method is characterized by comprising the following steps;