1000/1000
Hot
Most Recent
The accurate segmentation of lung nodules is challenging due to their small size, especially at the edge of the lung and near the blood vessels. Lung nodule segmentation is relatively broad and varies in terms of architecture, image pre-processing, and training strategy.
Table 1. Deep learning-based lung nodule segmentation architectures and their key information.
Study | Year | Architecture | Dataset | Approach | Performance |
---|---|---|---|---|---|
Pezzano et al. [13] | 2021 | CoLe-CNN | LIDC-IDRI | 2D Based U-Net Inception-v4 architecture Mean Square Error function |
F1 = 86.1 IoU = 76.6 |
Dong et al. [15] | 2020 | MV-SIR | LIDC-IDRI | 2D/3D Residual block Secondary input Multi views Voxel heterogeneity (VH) Shape heterogeneity (SH) |
ASD = 7.2 ± 3.3 HSD = 129.3 ± 53.3 DSC = 92.6 ± 3.5 PPV = 93.6 ± 2.2 SEN = 98.1 ± 11.3 |
Keetha et al. [2] | 2020 | U-DNet | LUNA16 | 2D Based U-Net Bi-FPN Efficient-Det Mish activity function |
DSC = 82.82 ± 11.71 SEN = 92.24 ± 14.14 PPV = 78.92 ± 17.52 |
Cao et al. [20] | 2020 | DB-ResNet | LIDC-IDRI | 2D/3D ResNet CIP Multiview Multiscale Central Intensity-Pooling |
DSC = 82.74 ± 10.19 ASD = 19 ± 21 SEN = 89.35 ± 11.79 PPV = 79.64 ± 13.34 |
Kumar el al. [3] | 2020 | V-Net | LUNA16 | 3D V-Net PReLU Only fully convolutional lays |
DSC = 96.15 |
Usman et al. [4] | 2020 | Adaptive ROI with Multi-view Residual Learning | LIDC-IDRI | 2D/3D the Deep Residual U-Net Adaptive ROI Multiview |
SEN = 91.62 PPV = 88.24 DSC = 87.55 |
Tang et al. [22] | 2019 | NoduleNet | LIDC-IDRI | 3D Multitask Residual-block Detection, FPR, segmentation Different loss function |
DSC = 83.10 CPM = 87.27 |
Huang et al. [5] | 2019 | Faster R-CNN | LUNA16 | 2D Faster RCNN Merge overlap FP reduction Based FCN |
ACC = 91.4 DSC = 79.3 |
Aresta et al. [6] | 2019 | iW-Net | LIDC-IDRI | 3D Based U-Net two points in the nodule boundary none heavy pre-processing steps augmentation |
IoU = 55 |
Hesamian et al. [23] | 2019 | Atrous convolution | LIDC-IDRI | 2D Atrous convolution Residual Network Weight loss Normalize to 0, 255 |
DSC = 81.24 Precision = 79.75 |
Liu et al. [24] | 2018 | Mask R-CNN | LIDC-IDRI | 2D Backbone: ResNet101, FPN transfer learning RPN FCN |
73.34 mAP 79.65 mAP |
Khosravan et al. [25] | 2018 | Semi-supervised multitask learning | LUNA16 | 3D Data augmentation Semi-supervised FP reduction |
SEN = 98 DSC = 91 |
Wu et al. [1] | 2018 | PN-SAMP | LIDC-IDRI | 3D 3D U-Net WW/WC Dice coefficient loss Segmentation, classification |
DSC = 73.98 |
Tong et al. [7] | 2018 | Improved U-NET network | LUNA16 | 2D U-Net Modify residual block Obtain lung parenchyma |
DSC = 73.6 |
Zhao et al. [8] | 2018 | 3D U-Net and Contextual Convolutional Neural Network | LIDC-IDRI | 3D 3D U-Net GAN Morphological methods Residual block Inception structure |
None |
Wang et al. [14] | 2017 | MV-CNN | LIDC-IDRI | 2D/3D Mutilview A multiscale patch strategy |
SEN = 83.72 PPV = 77.59 DSC = 77.67 |
Wang et al. [17] | 2017 | CF-CNN | LIDC-IDRI/GDGH | 2D/3D Central pooling 3D patch 2D views A sampling method Two datasets |
LIDC: DSC = 82.15 ± 10.76 SEN = 92.75 ± 12.83 PPV = 75.84 ± 13.14 GDGH: DSC = 80.02 ± 11.09 SEN = 83.19 ± 15.22 PPV = 79.30 ± 12.09 |