BDFNet: Boundary-Assisted and Discriminative Feature Extraction Network for COVID-19 Lung Infection Segmentation

Abstract

The coronavirus disease (COVID-19) pandemic has affected billions of lives around the world since its first outbreak in 2019. The computed tomography (CT) is a valuable tool for the COVID-19 associated clinical diagnosis, and deep learning has been extensively used to improve the analysis of CT images. However, owing to the limitation of the publicly available COVID-19 imaging datasets and the randomness and variability of the infected areas, it is challenging for the current segmentation methods to achieve satisfactory performance. In this paper, we propose a novel boundary-assisted and discriminative feature extraction network (BDFNet), which can be used to improve the accuracy of segmentation. We adopt the triplet attention (TA) module to extract the discriminative image representation, and the adaptive feature fusion (AFF) module to fuse the texture information and shape information. In addition to the channel and spatial dimensions that are mainly used in previous models, the cross channel-special context is also obtained in our model via the TA module. Moreover, fused hierarchical boundary information is integrated through the application of the AFF module. According to experiments conducted on two publicly accessible COVID-19 datasets, COVID-19-CT-Seg and CC-CCII, BDFNet performs better than most cutting-edge segmentation algorithms in six widely used segmentation metrics.

Appendix

In this section, we provide additional experiments and visualizations for the BDFNet model. Part A shows the visualization results of the heat map of the TA module and the AFF module. Part B compares the robustness differences between the TA module and other attention mechanisms. Part C shows the visual comparison of segmentation results between BDFNet and other advanced segmentation algorithms.

• A. The Visualization Results on TA Module and AFF Module

In Sect. 3.2, the ablation experiments on the TA module and the AFF module demonstrate the effectiveness of the proposed module. To further validate the proposed TA module and AFF module’s effectiveness intuitively, we randomly select some samples from COVID-19-CT-Seg dataset and use Grad-CAM [20] technology to visualize the gradients of segment prediction as a heat map, the bright area in the figure indicates the area participating in the segmentation.

Fig. 5.

The visualization results on TA module’s gradients through Grad-CAM.

Fig. 6.

The visualization results on AFF module’s gradients through Grad-CAM.

As shown in Fig. 5, compared with the ground truth map, the TA module is able to capture more accurate and comprehensive COVID-19 infection regions from CT images. In some cases, the network with the TA module can correctly identify the infected area that the original network misidentified.

Figure 6 shows the shape information captured by the network added with the AFF module. It can be seen from the figure that the obtained shape information basically depicts the outline of the infection area, which proves the AFF module’s capability to capture richer and more precise shape information from infection regions.

• B. The Compare of the Robustness for TA, PA, CA and IA

In Table 4 in Sect. 3.2, from the ablation experiment inside the TA module, it can be seen that although the TA module has achieved the best performance in multiple evaluation indicators, it does not lead too much. In order to verify the superiority of the proposed TA module, we performed a statistical analysis of the experimental results on the COVID-19-CT-Seg dataset, and the results are shown in Fig. 7.

Fig. 7.

The distribution of the results of different evaluation indicators of the internal ablation experiment in the TA module. The horizontal axis represents the network model that contains different attention mechanism modules, including TA module, PA module, CA module, IA module, and the vertical axis represents different evaluation indicators.

The fact is that the TA module achieves the best overall performance in COVID-19 infection regions segmentation, especially in terms of the metric of ${S}_{\alpha }$, ${E}_{\phi }$, and MAE. Figure 7 shows the evaluation distribution of the segmentation results of the four attention mechanism modules under dfferent evaluation metrics. From the figure, it can be seen that with the same evaluation parameters, the segmentation result obtained by the TA module is more stable. This shows that the TA module has a more robust segmentation ability for different test data. And it is not easy to be affected by extreme values in the segmentation results.

• C. Visual Comparison of Segmentation Results

In Sect. 3.3, Fig. 4 and Fig. 5 show that no networks can lead consistently on all evaluation metrics. Therefore, by visually comparing the segmentation results, we can more intuitively verify the pros and cons of the segmentation results.

Fig. 8.

Visual comparison of COVID-19 infected areas segmentation from different networks.

Figure 8 provide some comparison of COVID-19 infection regions segmentation results from different networks on COVID-19-CT-Seg dataset. As can be observed, the BDFNet yields segmentation results with more precise boundaries, especially in the subtle infected areas. From a medical point of view, it is very important to accurately distinguish the boundaries of the lesion area. It can help doctors determine the location and extent of the patient's infection and assist the doctor in diagnosis.