Fig 1.
Representative manually segmented IVUS images and mask images.
Upper images show IVUS images and lower images show mask images. Panel D is the mask image segmented from panel A. Similarly, panel E corresponds to panel B and panel F corresponds to panel C. The green, red, orange, and blue areas denote the lumen area, the medial plus plaque area, calcification, and a stent, respectively. The black areas show background.
Fig 2.
Illustration of U-Net used in this study.
The architecture of U-Net used in this study is illustrated. The blue block and orange block are two-dimensional convolutional layers (kernel size = 3 × 3, stride = 1) with dropout (rate = 0.1). The number in the blue block denotes the feature map size. The green block shows the maximum pooling layer (pool size = 2 × 2), the pink block shows the upsampling layer (size = 2 × 2), and the red dashed arrows show the concatenation of the two layers.
Fig 3.
“Normal” means that the images were without calcification or a stent. Of the 3738 IVUS images, 413 had both calcification and stents (purple double-headed arrow). In the test set, all 40 images with stents also had calcification. In the training set, 373 images with stents had calcification. AP = angina pectoris.
Fig 4.
Scatter plot showing a regression line for the lumen area in ground truth mask images and the lumen area in mask images predicted by our AI.
The light blue area shows the 95% confidence intervals of the regression line. Lumen areas of mask images predicted by our AI and lumen areas of ground truth mask images show a strong positive correlation (Spearman rank correlation).
Fig 5.
Representative images of cases successfully predicted by our AI.
The panels on the left (A, D, G) show IVUS images. The middle panels (B, E, H) show ground truth mask images that were manually segmented corresponding to IVUS images on the left in the same row. The panels on the right (C, F, I) show mask images predicted by our AI corresponding to IVUS images on the left in the same row. In the mask images, the green area shows the lumen area, the red area shows the medial plus plaque area, the orange area shows calcification, and the black area is background. Calcifications in the IVUS images were effectively segmented by our AI (F, I).
Fig 6.
Representative images of cases showing prediction failure by our AI.
The panels on the left (A, D, G) show IVUS images. The middle panels (B, E, H) show ground truth mask images that were manually segmented corresponding to IVUS images on the left in the same row. The panels on the right (C, F, I) show label images predicted by our AI corresponding to IVUS images on the left in the same row. In the mask images, the green area shows the lumen area, the red area shows the medial plus plaque area, the orange area shows calcification, the blue area shows a stent, and the black area is background. Two lumina were delineated in panel C because of misidentification of the coronary vein in panel A as the lumen of the coronary artery (blue arrowhead). The lumen edges in panel F were incorrectly delineated because of an artifact of the wire in panel D (pink arrowhead). The stent struts in panel I were misidentified as calcification (yellow arrowhead).
Table 1.
Results of the validation and test sets.
Table 2.
Results for stents and each component of the vessels in the test set.