Improved V-Net lung nodule segmentation method based on selective kernel

rong ze wang, jing ru meng, Fuchun Zhang

Abstract:

Segmentation algorithm based on deep learning has become the main method of pulmonary nodules segmentation; nevertheless, the accuracy and lightweight of most such models are difficult to coexist. In order to accurately segment lung nodules
in computed tomography images and make the model lightweight, this paper proposes a lightweight segmentation network
called SKV-Net, able to achieve good performance. The overall design of the network uses the original V-Net structure
and introduces a selective convolution kernel with soft attention in selective kernel networks to extract multi-scale feature
information. Adopting a suitable grouped convolution can effectively reduce the number of parameters in the model while
maintaining good segmentation performance. Experimental results indicate that the average segmentation accuracy of SKVNet is 1.3% higher than that of V-Net, and the number of parameters is only 42% those of V-Net. In this paper, the Luna16
public dataset of pulmonary nodules is used to test and evaluate the performance of various improved models. The results
suggest that the SKV-Net is superior to other models, achieving good segmentation performance and fast operation speed.
Moreover, the SKV-Net improves the segmentation of different types of pulmonary nodules. It has the advantages of high
precision and lightweight structure, which further indicate that it has significant clinical application value in the segmentation
task of pulmonary nodules.