Segmentation of Lung Nodules Using Improved 3D-UNet Neural Network

Zhitao Xiao, Bowen Liu, Lei Geng, Fang Zhang, Yanbei LiuSymmetry 2020 · DOI

TL;DR

This paper introduces a novel 3D convolutional neural network, 3D-Res2UNet, for segmenting lung nodules in CT scans. By integrating the multi-scale feature extraction capabilities of Res2Net into a 3D U-Net architecture, the model achieves state-of-the-art performance, significantly improving accuracy and recall for early lung cancer detection.

What Problem Does This Paper Address?

Lung cancer remains a leading cause of mortality worldwide, and early detection is crucial for effective treatment. Lung nodules are key indicators of early-stage lung cancer. However, accurately segmenting these nodules from CT scans is challenging due to their variable shapes, sizes, and the presence of complex surrounding tissues, which can lead to blurred edges and low contrast. Existing methods, whether traditional image processing or 2D deep learning approaches, often struggle with the 3D nature of CT data, leading to information loss and reduced accuracy. This paper aims to address these limitations by developing a more robust 3D segmentation method.

Methods

The core of this work is the proposed 3D-Res2UNet network. This architecture builds upon the successful U-Net framework, which is well-suited for semantic segmentation tasks. The key innovation lies in the integration of a 3D-Res2Net module within the U-Net's encoder path.

Here's a breakdown of the methodological advancements:

  • 3D Convolutional Neural Network: Recognizing that CT scans are inherently 3D, the authors employ 3D convolutional layers to process the volumetric data, preserving spatial information that 2D methods would lose.
  • Res2Net Module: This module is adapted into a 3D version (3D-Res2Net). Res2Net is known for its ability to extract multi-scale features with finer granularity by creating hierarchical connections within residual blocks. This helps the network capture features at different scales and improves its ability to handle variations in nodule size and shape.
  • Integration into 3D U-Net: The 3D-Res2Net module is strategically placed within the down-sampling stages of the 3D U-Net. This placement allows the network to progressively extract richer, multi-scale features. The U-Net's encoder-decoder structure then uses these features for segmentation, with the decoder part up-sampling and combining features to reconstruct the segmentation mask.
  • Addressing Gradient Issues: The residual learning aspect of ResNet helps mitigate gradient disappearance and explosion problems, which are common in deep networks, leading to more stable training and improved accuracy.

Datasets

The proposed method was trained and evaluated on the LUNA16 (Lung Nodule Analysis 2016) public dataset. This dataset comprises 888 CT image sets containing a total of 1186 lung nodules. The data was split into training (70%), testing (20%), and validation (10%) sets. A nodule was considered valid if at least three out of four radiologists agreed on its presence and its radius was greater than 3 mm.

Key Insights and Ideas

  • The Power of 3D for Medical Imaging: The paper strongly reinforces the idea that for volumetric medical data like CT scans, 3D convolutional networks are essential for preserving spatial context and achieving superior performance compared to 2D approaches.
  • Multi-Scale Feature Extraction is Crucial: Lung nodules are highly variable. The success of the 3D-Res2Net module highlights the importance of explicitly designing networks that can capture features at multiple scales simultaneously, enabling better discrimination of nodules from surrounding structures.
  • Strategic Module Placement Matters: The experimental results showing that placing the 3D-Res2Net module in specific down-sampling layers (second and third) yields the best performance suggests that the architecture's design and the flow of information are critical for optimal results.
  • Balancing Accuracy and Computational Cost: While deeper networks can improve accuracy, they also increase computational burden. The authors acknowledge this trade-off, suggesting that their approach strikes a good balance, particularly for small nodule detection where extreme precision might not always be practical.

Key Results

The 3D-Res2UNet model demonstrated impressive performance on the LUNA16 dataset:

  • Dice Coefficient: Achieved a high score of 95.30%, indicating excellent overlap between the predicted and actual nodule segmentations. This is a significant improvement over other tested methods.
  • Recall Rate: Reached 99.1%, meaning the model is highly effective at identifying almost all actual lung nodules present in the scans.
  • False Positives: The model reported 276.3 false positive nodules per CT sample. While this is a notable number, it's presented in the context of improving segmentation fit and recall, which are critical for assisting radiologists.
  • Model Comparison: The 3D-Res2UNet outperformed other networks, including standard UNet, 3D-UNet, and 3D-UNet with fully connected CRFs, in terms of the Dice coefficient.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • High Training Costs: The use of 3D convolutions and a deeper network architecture leads to significant computational requirements for training.
  • False Positive Rate: While recall is high, the number of false positives remains a challenge. Future work could focus on further reducing false positives, perhaps by incorporating more sophisticated post-processing steps or refining the network's ability to distinguish nodules from similar-looking structures like blood vessels.
  • Handling Irregular Nodules: The paper notes that nodules with rough or jagged edges, which deviate from typical nodule shapes, can be challenging to segment accurately. This suggests a need for network architectures that are even more robust to extreme variations.
  • Generalizability: While the method shows promise for small particle detection in other fields, further validation on diverse datasets and different types of medical imaging would be beneficial.

Summary updated

BibTeX
@article{xiao2020segmentation,
  author = {Zhitao Xiao and Bowen Liu and Lei Geng and Fang Zhang and Yanbei Liu},
  title = {{Segmentation of Lung Nodules Using Improved 3D-UNet Neural Network}},
  journal = {Symmetry},
  year = {2020},
  doi = {10.3390/sym12111787},
}
Figure