MSD-Net: Multi-Scale Discriminative Network for COVID-19 Lung Infection Segmentation on CT

Bingbing Zheng, Yaoqi Liu, Yu Zhu, Fuli Yu, Tianjiao Jiang, Dawei Yang, Tao XuIEEE Access 2020 · DOI

TL;DR

This paper introduces MSD-Net, a novel deep learning model designed for accurate multi-class segmentation of COVID-19 lung infections in CT scans. By incorporating multi-scale feature extraction and attention mechanisms, MSD-Net demonstrates superior performance in identifying and delineating ground-glass opacities, interstitial infiltrates, and consolidation, offering a valuable tool for clinical diagnosis and treatment.

What Problem Does This Paper Address?

The COVID-19 pandemic has highlighted the urgent need for rapid and accurate diagnostic tools. While RT-PCR tests can confirm infection, they don't quantify disease severity. Computed Tomography (CT) scans offer a fast, non-invasive method to visualize lung abnormalities. However, segmenting the different types of COVID-19 lung infections (ground-glass opacities, interstitial infiltrates, and consolidation) is challenging due to their varied appearances, sizes, and frequent co-occurrence within the same CT image. This complexity makes manual segmentation time-consuming and prone to inter-observer variability, necessitating an automated and precise solution to aid clinicians in diagnosis and treatment planning.

Methods

The core of this work is the MSD-Net (Multi-Scale Discriminative Network), a deep learning architecture built upon the U-Net framework. Key novel components include:

  • Pyramid Convolution Block (PCB): This block enhances the network's ability to capture features at multiple scales by using kernels of varying sizes and numbers. This is crucial for segmenting infected areas that can range significantly in size.
  • Channel Attention Block (CAB): This attention module fuses features from adjacent encoder-decoder stages, focusing on relevant areas and improving the discriminative power of the features. It's designed to highlight areas of interest for segmentation.
  • Residual Refinement Block (RRB): Inspired by residual blocks, this component further refines feature maps, strengthening the discriminative ability of each stage within the network.

The network utilizes a ResNet-101 as its backbone encoder and incorporates a global average pooling layer to capture global semantic context. The overall architecture processes CT images through multiple stages, with each stage benefiting from the PCB, CAB, and RRB modules before producing a final segmentation output.

Datasets

The study utilized a COVID-19 CT segmentation dataset collected from The Affiliated Hospital of Qingdao University. This dataset comprises CT sequences from 18 COVID-19 patients and 18 control subjects. A total of 4780 2D CT axial slices were extracted, with 2506 slices containing infected lesions. These slices were annotated by professional doctors to identify three categories of lung infections: ground-glass opacities, interstitial infiltrates, and consolidation. The dataset was split into training (3824 images) and validation (956 images) sets using 5-fold cross-validation.

Key Insights and Ideas

  • Multi-scale segmentation is critical: The diverse sizes and appearances of COVID-19 lung lesions necessitate a network capable of capturing information across multiple scales, which the PCB module effectively addresses.
  • Attention mechanisms enhance discriminability: The CAB module plays a significant role in improving segmentation accuracy by allowing the network to focus on relevant features and suppress irrelevant background information.
  • Refinement is key for accuracy: The RRB module helps to further refine feature representations, leading to more precise segmentation boundaries.
  • Focal Loss for Imbalanced Data: The authors employed Focal Loss to mitigate the challenge of class imbalance, particularly the large proportion of background pixels, thereby improving the network's sensitivity to under-represented infected regions.

Key Results

The proposed MSD-Net achieved state-of-the-art performance on the COVID-19 CT segmentation dataset.

  • Dice Similarity Coefficient (DSC): The model achieved DSC scores of 0.7422 for ground-glass opacities, 0.7384 for interstitial infiltrates, and 0.8769 for consolidation. These represent significant improvements over baseline models like U-Net and even attention-based variants.
  • Sensitivity and Specificity: MSD-Net demonstrated strong sensitivity and specificity across all categories, with results for the three infection types being (0.8593, 0.9742), (0.8268, 0.9869), and (0.8645, 0.9889) respectively.
  • Robustness: The model showed good performance even when CT data was corrupted with Gaussian noise, indicating its robustness to artifacts.
  • Qualitative Improvements: Visual comparisons in Figure 4 clearly show MSD-Net producing more accurate and complete segmentations compared to other methods, closely matching the ground truth.

Limitations and Future Work

The authors acknowledge several limitations:

  • Insufficient Samples for Mild Symptoms: The dataset contains fewer samples of mild COVID-19 symptoms, which are inherently vague and difficult to distinguish. While Focal Loss helps, further optimization is needed.
  • Dataset Size: While efforts were made to mitigate overfitting (e.g., data augmentation), the dataset size could still be a limiting factor.

Future work suggested includes:

  • Stronger Backbone Models: Exploring more powerful backbone architectures.
  • Optimal Loss Functions: Investigating alternative or improved loss functions.
  • New Data Enhancement Methods: Developing advanced techniques for data augmentation to further improve accuracy.

Summary updated

BibTeX
@article{zheng2020msd,
  author = {Bingbing Zheng and Yaoqi Liu and Yu Zhu and Fuli Yu and Tianjiao Jiang and Dawei Yang and Tao Xu},
  title = {{MSD-Net: Multi-Scale Discriminative Network for COVID-19 Lung Infection Segmentation on CT}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3027738},
}
Figure