Abstract:An important prerequisite for improving the reliability of lung cancer surveillance and clinical interventions is accurate lung nodule segmentation. Although deep learning is effective at performing medical image segmentation, lung CT image heterogeneity, nodule size, shape, and location variations, convolutional localized feature extraction characteristics, the receptive field limitations of continuous downsampling, lesion edge information losses, fuzzy boundary segmentation challenges, and the low segmentation accuracy achieved when segmenting lung CT images using deep learning remain. An edge-enhanced multiscale Sobel coordinate attention-atrous spatial convolutional pooling pyramid V-Net (SCA-VNet) algorithm for lung nodule segmentation was proposed to solve these problems. First, a residual edge enhancement module was designed, which was used to enhance the edges of the original data. Using an edge detection operator in combination with a residual module, this module could reduce data redundancy and alleviate the gray level similarity between the foreground and background. Then, a 3D atrous spatial convolutional pooling pyramid module set different expansion rates, which could obtain feature maps under different receptive fields and capture the multiscale information of the segmentation target. Finally, a three-dimensional coordinate attention network (3D CA-Net) module was added to the encoding and decoding paths to extract channel weights from multiple dimensions. This step propagated the spatial information in the coding layer to the subsequent layers, and it could reduce the loss of information during the forward propagation process. The proposed method achieved a Dice coefficient of 87.50% on the lung image database consortium and image database resource initiative (LIDC-IDRI). It significantly outperformed the existing lung nodule segmentation models (UGS-Net, REMU-Net, and multitask models) and compared favorably with the Med3D, CENet, and PCAM_Net segmentation models in terms of their Dice coefficients, which were 3.37%, 2.2%, and 1.43%, respectively. The experimental results showed that the proposed SCA-VNet model attained improved lung nodule segmentation accuracy and laid a good foundation for improving the early detection rate of lung cancer.

SaraNet: Semantic Aggregation Reverse Attention Network for Pulmonary Nodule Segmentation

ResDSda_U-Net: A Novel U-Net-Based Residual Network for Segmentation of Pulmonary Nodules in Lung CT Images

LUNG NODULE SEMANTIC SEGMENTATION WITH BI-DIRECTION FEATURES USING U-INET

NoduleNet: Decoupled False Positive Reductionfor Pulmonary Nodule Detection and Segmentation

S-Net: an S-shaped network for nodule detection in 3D CT images

An Intelligent Model For Benign And Malignant Pulmonary Nodule Analysis Using U-Net Networks And Multilevel Attention Mechanisms

ECANodule: Accurate Pulmonary Nodule Detection and Segmentation with Efficient Channel Attention.

An improved V-Net lung nodule segmentation model based on pixel threshold separation and attention mechanism

NoduleNet: Decoupled False Positive Reduction for Pulmonary Nodule Detection and Segmentation

Multiscale lung nodule segmentation based on 3D coordinate attention and edge enhancement

DAG‐Net: Dual‐Branch Attention‐Guided Network for Multi‐Scale Information Fusion in Lung Nodule Segmentation

Automatic Segmentation of Lung Noudles Using Improved U-Net NetWork

AWEU-Net: An Attention-Aware Weight Excitation U-Net for Lung Nodule Segmentation

SM-GRSNet: sparse mapping-based graph representation segmentation network for honeycomb lung lesion

Lung Nodule Segmentation and Recognition Algorithm Based on Multiposition U-Net

An Efficient DA-Net Architecture for Lung Nodule Segmentation

Joint Learning for Pulmonary Nodule Segmentation, Attributes and Malignancy Prediction

LSSANet: A Long Short Slice-Aware Network for Pulmonary Nodule Detection

U-Det: A Modified U-Net architecture with bidirectional feature network for lung nodule segmentation

Pulmonary nodule segmentation based on REMU-Net

Nodule-Plus R-CNN and Deep Self-Paced Active Learning for 3D Instance Segmentation of Pulmonary Nodules