A Multi-Scale Liver Tumor Segmentation Method Based on Residual and Hybrid Attention Enhanced Network with Contextual Integration

Liyan Sun, Linqing Jiang, Mingcong Wang, Zhenyan Wang, Yi Xin2024 · DOI

TL;DR

This paper introduces RHEU-Net, a novel deep learning architecture for segmenting liver tumors in CT scans. By integrating residual modules, a hybrid attention mechanism, and multi-scale feature enhancement, RHEU-Net significantly improves accuracy in identifying both liver and tumor regions, addressing limitations of existing methods in handling complex tumor shapes and low contrast.

What Problem Does This Paper Address?

Accurate segmentation of liver tumors from CT scans is crucial for timely diagnosis and effective treatment of liver cancer, a disease with high mortality rates. Traditional segmentation methods struggle with the inherent challenges of medical imaging, such as blurred boundaries, low contrast between tumors and surrounding tissues, and significant variations in tumor size and shape. While U-Net and its variants have shown promise, they often fall short in capturing fine details, adapting to diverse tumor morphologies, and integrating contextual information effectively. This leads to issues like over-segmentation and under-segmentation, particularly for smaller tumors. This paper aims to bridge this gap by proposing a more robust and accurate segmentation method.

Methods

The proposed RHEU-Net builds upon the U-Net architecture with several key enhancements:

  • Residual Modules: Standard convolutional blocks in the encoder and decoder are replaced with improved residual modules. This helps preserve low-level features and mitigates the vanishing gradient problem, leading to more stable training and better feature transfer.
  • Hybrid Gated Attention (HGA) Module: Integrated before skip connections, HGA processes channel and spatial attentions simultaneously. It uses a gating mechanism to dynamically adjust the fusion strategy of these attentions, allowing the model to focus on critical features and better capture multi-dimensional and irregular structures.
  • Multi-Scale Feature Enhancement (MSFE) Layer: This layer replaces the bottleneck block and utilizes atrous convolutions with different dilation rates. This expands the receptive field, enriching semantic information and capturing global contextual information, which is vital for segmenting tumors of varying sizes and compensating for lost spatial information at the bottleneck.
  • BCEDiceLoss: A combined loss function that synergistically leverages Binary Cross-Entropy (BCE) for pixel-level accuracy and Dice loss for assessing overlap, aiming for a balance between precision and recall.

Datasets

The study utilizes the publicly accessible LiTS2017 dataset. This dataset comprises 201 abdominal enhanced CT scans from six medical institutions, with 131 scans annotated for liver and liver tumors. The dataset contains 58,638 CT slices, each with a resolution of 512x512 pixels. For this research, 101 annotated scans were used for preprocessing and training, and 30 were randomly selected for the test set.

Key Insights and Ideas

  • Synergy of Attention: The paper highlights that combining spatial and channel attention mechanisms in a gated manner (HGA) is more effective than sequential or independent application, allowing for dynamic feature fusion.
  • Multi-Scale is Crucial for Tumors: The MSFE layer's ability to capture features at various scales is shown to be particularly important for the diverse sizes and shapes of liver tumors, addressing a key limitation of single-scale approaches.
  • Residual Connections for Stability: The integration of residual modules not only aids in feature extraction but also significantly improves training stability by combating vanishing gradients, a common issue in deep networks.
  • Holistic Integration Outperforms Individual Modules: The ablation studies demonstrate that while each proposed module (Residual, HGA, MSFE) offers improvements, their combined synergistic effect in RHEU-Net yields the most substantial gains in segmentation performance.

Key Results

  • Liver Segmentation: RHEU-Net achieved a Dice coefficient of 95.72%, Jaccard index of 91.49%, and Precision of 96.23% on the LiTS2017 dataset. This represents a significant improvement over the baseline U-Net and other state-of-the-art models.
  • Tumor Segmentation: For tumor segmentation, RHEU-Net achieved a Dice coefficient of 70.19%, Jaccard index of 61.26%, and Precision of 85.45%. This is a substantial improvement, particularly for smaller and more challenging tumors.
  • Ablation Studies: The proposed RHEU-Net model, integrating all three key modules, showed a 1.45% increase in Dice score for liver segmentation and a 3.36% increase for tumor segmentation compared to the baseline U-Net.
  • Qualitative Results: Visual comparisons in Figure 11 show RHEU-Net producing segmentation masks that more closely align with the gold standard, accurately capturing smaller tumors and finer liver contours where other models struggled with over- or under-segmentation.

Limitations and Future Work

  • Data Acquisition Challenge: The authors acknowledge that acquiring medical data remains a significant challenge.
  • Future Directions:
    • Developing more efficient lightweight network architectures to keep pace with evolving medical technology.
  • Exploring RHEU-Net's potential for expanding CT imaging workflows beyond liver segmentation.
  • Optimizing RHEU-Net for processing sensor-collected data from various organs and diagnostic requirements.

Summary updated

BibTeX
@misc{sun2024multi,
  author = {Liyan Sun and Linqing Jiang and Mingcong Wang and Zhenyan Wang and Yi Xin},
  title = {{A Multi-Scale Liver Tumor Segmentation Method Based on Residual and Hybrid Attention Enhanced Network with Contextual Integration}},
  year = {2024},
  doi = {10.3390/s24175845},
}
Figure