Table 1.
Summary of existing studies.
Fig 1.
Shows the workflow of lung segmentation: In (a) we show an original image of lung CT scan, and (b) shows a binary mask of the original image. Since the pixel intensities of GGO and bone regions are similar to each other, the binary mask shows an incomplete structure of the lung. Some parts of the lungs are missing. To restore the missing parts, we applied the convex hull algorithm to the binary mask. The convex hull algorithm detects convex points [red dots in Fig 1C], and we compute pairwise Euclidean distances to connect nearest-neighbor points by a green line if the distance between them is less than 80 pixels [Fig 1C]. Fig 1D and 1E show the restored mask and the corresponding RGB image, respectively.
Fig 2.
Shows the workflow diagram of the point cloud and PointNet++ for pattern analysis of GGO distribution.
The PointNet++ architecture used in this study is the same as the original architecture in [28 and 29]. The data preparation pipeline is new, which we developed for generating point clouds from 3D lung CT scans. We stacked all segmented lung images to construct a 3D image and used sampling and grouping to select 2048 points for point clouds. We fed those points to the PointNet++ architecture, which was constructed with a multiple set of abstraction layers. Each set abstraction layer consisted of a sampling layer, grouping layer, and PointNet [28]. Multiple set abstraction layers formed a hierarchical feature learning architecture, which was divided into two parts: Segmentation and classification. For data segmentation, we merged local and global features to get the score of individual data points. For classification, global features were fed to two fully connected (FC) layers. MLP—Multilayer Perceptron.
Fig 3.
Shows the PointNet++ training and testing curves.
Training and testing accuracies increase with increasing wall time and concomitantly the loss decreases with increasing wall time. This is an ideal behavior of a properly trained model.
Fig 4.
Final results after training the PointNet++ using point cloud data are shown here.
The 3D images show different regions from different angles. Cyan color points/bubbles represent non-GGO regions and green color points/bubbles represent GGO regions.
Table 2.
Performance results of PointNet++.
Fig 5.
(a) The architecture of the CNN for abnormality prediction. We used the segmented lungs and GGOs as input images (described in section 3.1). (b) Workflow diagram for abnormality prediction. We used the CNN architecture mentioned in Fig 5A. The output of the CNN network was fed to the Cox model to compute abnormality. The output of abnormality scores is represented by a heatmap. Orange to red color denotes high-abnormality regions.
Fig 6.
GGO shape analysis of COVID-19 data using Minkowski tensors.
(a) Probability distribution shows that GGOs for all the images lie between -800 and -150 HU with peaks in the distribution centered around -600 HU. (b) Radial distribution function versus distance shows a peak around 125 pixels.