Segmentation of Pulmonary Nodules Based on IAD-VNet

Pulmonary nodules are lung cancer's early manifestations. In order to judge benign and malignant diagnoses, it is necessary to manually mark the nodule position on CT slices, which is time-consuming and unstable. In order to accurately segment pulmonary nodules in CT images, an IAD-VNet network...

Full description

Saved in:
Bibliographic Details
Main Authors: Wu, Hai-Yan, Wei, Gong-Zheng, Li, Yuan-Ye
Format: Conference Proceeding
Language:English
Subjects:
Computed tomography
Dense connections
Image segmentation
Iterative deep aggregation
LUNA16
Lung
Lung cancer
Lung nodule segmentation
Medical services
Shape
Three-dimensional displays
V-Net
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Pulmonary nodules are lung cancer's early manifestations. In order to judge benign and malignant diagnoses, it is necessary to manually mark the nodule position on CT slices, which is time-consuming and unstable. In order to accurately segment pulmonary nodules in CT images, an IAD-VNet network is proposed to automatic segmentation of pulmonary nodules and to assist doctors in diagnosis. In IAD-VNet, a deep aggregation structure is introduced, and the IAD-VNet network iteratively aggregates feature maps of different scales to perform deep fusion of shallow and deep features, which strengthens the transfer of features at different depths and refines shallow features fully. A deep supervision strategy is introduced to ensure that the network is fully trained. At the same time, the weight learning module is utilized to obtain the weights of different depth features, and a weighted summation is performed to obtain the final segmentation result. Then the segmentation accuracy is improved. Furthermore, the IAD-VNet network introduces dense connections in some deep encoding/decoding modules, which improves the utilization of the feature map by the network and alleviates the gradient disappearance of the network. The results on the LUNA16 dataset show that the dice similarity coefficient, precision, and recall achieved by the segmentation model are 0.835, 0.846, and 0.837, respectively. Compared with other segmentation methods, the IAD-VNet network has good robustness and can obtain good segmentation results for lung nodules of different shapes and sizes, providing a good basis for further diagnosis.
ISSN:2688-0938
DOI:10.1109/CAC57257.2022.10055019