PaperlayerView exact revision on arXiv

Source-linked AI summary

Dual-branch residual network for lung nodule segmentation

Haichao Cao, Hong Liu, Enmin Song, Chih-Cheng Hung, Guangzhi Ma, Xiangyang Xu, Renchao Jin, Jianguo Lu

arXiv:1905.08413v1cs.CVeess.IV

TL;DR

Heterogeneous lung nodules can resemble their surroundings, making accurate segmentation difficult even though it matters for subsequent analysis. DB-ResNet combines multi-view and multi-scale features, center-voxel intensity features, and boundary-based weighted sampling; it reports an average Dice score of 82.74% on LIDC.

  • Problem

    Nodule heterogeneity and similarity to surrounding tissue make accurate segmentation difficult, although segmentation matters for subsequent analysis.

    Source: §1 · p. 2

  • Method

    DB-ResNet combines multi-view and multi-scale voxel features with center-voxel intensity features and boundary-based weighted sampling.

    Source: §1 · p. 3

  • Results

    82.74% average Dice score on the LIDC dataset; DB-ResNet also successfully segmented challenging juxtapleural and small nodules.

    Source: §6 · p. 19

  • Takeaways & Limitations

    The reported results include successful segmentation of challenging juxtapleural and small nodules.

    Source: §6 · p. 19

Abstract

from arXiv · show

An accurate segmentation of lung nodules in computed tomography (CT) images is critical to lung cancer analysis and diagnosis. However, due to the variety of lung nodules and the similarity of visual characteristics between nodules and their surroundings, a robust segmentation of nodules becomes a challenging problem. In this study, we propose the Dual-branch Residual Network (DB-ResNet) which is a data-driven model. Our approach integrates two new schemes to improve the generalization capability of the model: 1) the proposed model can simultaneously capture multi-view and multi-scale features of different nodules in CT images; 2) we combine the features of the intensity and the convolution neural networks (CNN). We propose a pooling method, called the central intensity-pooling layer (CIP), to extract the intensity features of the center voxel of the block, and then use the CNN to obtain the convolutional features of the center voxel of the block. In addition, we designed a weighted sampling strategy based on the boundary of nodules for the selection of those voxels using the weighting score, to increase the accuracy of the model. The proposed method has been extensively evaluated on the LIDC dataset containing 986 nodules. Experimental results show that the DB-ResNet achieves superior segmentation performance with an average dice score of 82.74% on the dataset. Moreover, we compared our results with those of four radiologists on the same dataset. The comparison showed that our average dice score was 0.49% higher than that of human experts. This proves that our proposed method is as good as the experienced radiologist.

1. Introduction

The introduction frames lung-nodule segmentation as difficult because nodule appearance varies and can resemble surrounding tissue, while accurate masks matter for subsequent analysis. It introduces DB-ResNet and its multi-view, multi-scale, intensity-feature, and weighted-sampling components.

1. Introduction

  • Accurate nodule segmentation matters because it can affect subsequent analysis; manual treatment is tedious and can increase diagnostic differences among doctors.Source: §1 · p. 2
  • Nodule heterogeneity and visual similarity to surrounding tissue make accurate segmentation difficult across juxtapleural, ground-glass, calcific, cavitary, and small nodules.Juxtapleural nodules can resemble the lung wall, ground-glass nodules have low contrast, and small nodules can resemble surrounding noise.Source: §1 · p. 2
  • DB-ResNet uses multi-view branches for upper, middle, and lower slices and multi-scale branches for three scales of the middle slice.The extracted features are used to classify each voxel.Source: §1 · p. 3
  • The central intensity-pooling layer preserves intensity features centered on the target voxel rather than boundary intensity information, incorporating traditional intensity features into the CNN.Source: §1 · p. 3
  • The improved weighted sampling strategy selects voxels using nodule-boundary counts so that small nodules can be adequately sampled.Source: §1 · p. 3

2. Related Work

Prior lung-nodule segmentation work spans morphology, region growing, energy optimization, and machine learning. The cited approaches use different image cues and assumptions, with reported limitations for irregular, juxtapleural, ground-glass, or small nodules.

2. Related Work

  • Morphological methods isolate nodules using operations and connected regions, while other variants address lung-wall separation and juxtapleural nodules.The review describes morphology-based nodule segmentation as challenging.Source: §2 · p. 4
  • Region growing requires a seed and stopping condition; cited methods are limited for juxtapleural or irregularly shaped nodules,.The review says basic region growing is best adapted to calcified nodules; Dehmeshki et al. added intensity, distance, fuzzy connectivity, and peripheral contrast.Source: §2 · p. 4
  • Energy-optimization methods use level sets [22–25], shape priors, or graph cuts, but are reported as poorly adapted to ground-glass and juxtapleural nodules.Source: §2 · p. 4
  • Machine-learning approaches design and extract features for voxel classification, including invariant, texture, shape, and vascular features.Source: §2 · pp. 4–5
  • CNN methods classify voxels, and have used multiple views or branches; the cited central-focused CNN was not ideal for small nodules.The review also notes FCN approaches including 2D and 3D U-Net.Source: §2 · p. 5

3. Our Proposed Method

DB-ResNet combines multi-view and multi-scale CT information with center-voxel intensity features, boundary-weighted sampling, and post-processing to segment lung nodules.

3.1.1. Network structure

  • The network has two deep branches with the same structure but different inputs, whose features are concatenated before a shared fully connected layer.Residual blocks and bottleneck convolutions are used to reduce network parameters and increase network depth.Source: §3.1.1 · p. 7, §3.1.2 · p. 8

3.1.2. Dual-Branch Architecture

  • The dual-branch architecture captures features from three contiguous slices and from three patch scales centered on the target voxel.The multi-view branch uses the previous, current, and next slices; the multi-scale branch uses 65×65, 50×50, and 35×35 patches rescaled to 35×35.Source: §3.1 · p. 5, §3.1.2 · p. 8

3.1.3. The Central Intensity-Pooling

  • Central intensity-pooling extracts intensity features centered on the target voxel to complement the CNN features.The method uses local and global pooling kernels of 1×1 and 3×3, respectively.Source: §3.1.1 · p. 7, §3.1.3 · p. 9

3.2. The weighted sampling strategy

  • The improved sampling strategy addresses inadequate small-nodule sampling by selecting twice as many nodule samples as boundary voxels and matching the non-nodule sample count.The method takes all voxels for nodules under 6 mm when there are too few to meet that target.Source: §3.2 · p. 10
  • 78.89% to 80.30% average dice score after the improved boundary-based sampling strategy.For nodules under 6 mm, the method samples all nodule voxels if the target of twice the boundary-voxel count exceeds the total.Source: §3.2 · p. 10

3.3. Post-processing

  • For semi-automatic segmentation, a manually specified bounding box on one starting slice is propagated across slices, with connected-region selection used to remove noise.Propagation stops when the intersection area with the previous slice is less than 30% of the previous slice's nodule area.Source: §3.3 · pp. 10–11

4. Data and Experiments

The experiments use LIDC nodules divided into training, validation, and test sets, evaluate segmentation with overlap and surface-distance criteria, and include implementation details and an ablation study.

4.1. Data

  • 986 LIDC nodules annotated by four radiologists were split into training, validation, and testing sets of 387, 55, and 544 nodules.Ground-truth boundaries used the 50% consistency criterion.Source: §4.1 · p. 11

4.2. Evaluation criteria

  • DSC and average surface distance are the primary segmentation metrics, with positive predictive value and sensitivity as auxiliary measures.DSC measures overlap between two segmentation results.Source: §4.2 · p. 12

4.3. The Detail of Implementation

  • Training used 0.47 million voxel patches, early stopping, a maximum of 20 epochs, and took about 31 hours to converge.The training patches were extracted from the LIDC training set using weighted sampling.Source: §4.3 · pp. 12–13

4.4. Ablation Study

  • The ablation study evaluates DB-ResNet components on the LIDC test set, including scale, boundary-weighted sampling, dual-branch architecture, ResNet, CIP placement, and post-processing.The table caption identifies the tested components but does not provide their results.Source: Table 3 · p. 13, §4.4 · p. 13

DSC ASD SEN PPV

The ablations report gains from boundary-based weighted sampling and residual blocks, while central intensity-pooling performs best in one configuration but declines when intensity features dominate. Post-processing yields smaller improvements across the four measures.

Effect of Boundary-based Weighted Sampling (BWS)

Effect of Residual Network

  • DSC 82.37% for DB-ResNet32 was a two-percent increase over the preceding 80.30% configuration.The residual-network variant replaced two convolution blocks with two residual blocks.Source: （2）Effect of Residual Network · p. 13

Effect of Central Intensity-Pooling

Effect of Post-processing

  • Post-processing improved all four performance measures, although the improvement was not significant.The comparison was between the ninth and last rows of Table 3.Source: （4）Effect of Post-processing · p. 14

5. Results and Discussion

DB-ResNet is reported to outperform existing segmentation methods and to compare favorably with radiologists on the LIDC testing set. Its results are reported as robust across nodule characteristics and challenging sizes, with visual comparisons highlighting performance on several nodule types.

5.1. Overall performance

  • In the testing set, most nodules had a DSC value higher than 0.8, based on the plotted DSC-versus-nodule-count histogram; Fig. 4 presents the LIDC testing-set DSC distributions.Source: §5.1 · p. 14, Figure 4 · p. 15

5.2. Robustness of Segmentation

  • DB-ResNet achieved similar performance across nodule-characteristic groups and was reported to handle attached, non-attached, small, and larger nodules robustly.The robustness analyses are summarized in Tables 5 and 6.Source: §5.2 · pp. 15–16, Table 5 · p. 16, Table 6 · p. 16

5.3. Experimental Comparison

  • DB-ResNet was superior to existing segmentation methods under comparison conditions that included boundary-based sampling, central intensity-pooling, and post-processing.For methods using the same architecture, DB-ResNet performed best; the visual comparison also reports robust segmentation across representative nodule types.Source: §5.3 · pp. 17–18, Table 7 · p. 17, Table 8 · p. 17
  • For juxtapleural and small nodules, DB-ResNet segmentations had large overlap with the ground-truth contours across the displayed slices.Fig. 6 identifies the small example as 4.8 mm in diameter and distinguishes DB-ResNet contours from ground truth.Source: §5.3 · p. 18, Figure 6 · p. 19

6. Conclusion

The paper concludes that DB-ResNet provides accurate lung nodule segmentation, including for juxtapleural and small nodules. A fully automated system combining it with a DSSD-based detector is proposed as future work.

6. Conclusion

  • DB-ResNet achieved an average Dice score of 82.74% on the LIDC dataset.The conclusion describes the model as using dual-branch feature extraction.Source: §6 · p. 19
  • The model was reported to successfully segment challenging juxtapleural nodules and small nodules.Source: §6 · p. 19
  • Future work proposes combining DB-ResNet with a DSSD-based detector to create a fully automated lung-nodule segmentation system.Source: §6 · p. 19
Loading 1905.08413v1…