A 3D Semantic Segmentation Method of Lung Parenchyma in CT Images Based on Deep Neural Network

Abstract

The invention discloses a three-dimensional semantic segmentation method of lung parenchyma in CT images based on a deep neural network. The segmentation method includes an offline part and an online part. Offline part: including 4 steps: preprocessing of data sets; construction of full convolutional neural network framework; construction of GRU loop convolutional neural network framework; network training. Online part: including 5 steps: image preprocessing; pixel feature extraction and fusion; voxel feature extraction and fusion; segmentation output; segmentation result evaluation. Feature fusion includes feature fusion between encoding layer and decoding layer, feature fusion between anatomical structure layer and layer. By designing a deep neural network model with gated recurrent units, using the prior information of the anatomical structure of the lungs, spatial features are extracted to effectively characterize the apparent evolution relationship between tomographic sequences, and to achieve accurate lung parenchyma in CT images. 3D semantic segmentation.

Description

technical field [0001] The invention belongs to the image segmentation technology in the field of medical image processing, in particular to a three-dimensional semantic segmentation method of lung parenchyma in CT images based on deep neural networks. Background technique [0002] Lung disease has always been a serious threat to human health. Lung cancer has the highest mortality rate of all malignant tumors, and the overall 5-year survival rate does not exceed 15%. With the rapid development of my country's economic level, people's attention to their own health is also increasing. Early diagnosis and early treatment of diseases have increasingly become the common concern of the whole society. In order to reduce the workload of doctors and detect lung diseases more quickly and accurately, the application of medical image processing technology in the auxiliary diagnosis of lung diseases is of great significance to the auxiliary diagnosis of clinical lung diseases. The segme...

AI Technical Summary

Problems solved by technology

The shortcomings of the region growing method are: (1) how to define the regional consistency criterion; (2) the segmentation result has a great relationship with the selection of the seed point; (3) this method is also very sensitive to noise and may Formation of porous or even discontinuous regions at all

[0008] Although these 2D CNN-based methods greatly improve the segmentation accuracy of traditional handcrafted feature-based methods, for volumetric medical image analysis, the effect may not be optimal because they cannot fully utilize the special information in the volumetric data.

Later, researchers proposed a 2.5D method to add richer spatial information, but it is still limited by the 2D kernel

Images