Predicting CT-Based Coronary Artery Disease Using Vascular Biomarkers Derived from Fundus Photographs with a Graph Convolutional Neural Network

Figures

• 

  Figure 1. The pipeline diagram for this study. Patients were recruited for CCTA and fundus eye examinations on the same day. The CCTA scans were diagnosed and stratified based on the CAD-RADS guideline. (A,B): The fundus images were acquired and pre-processed by brightness normalization. (C,E): The blood vessel centerlines were manually annotated, and the bifurcation relationship at each junction was labeled. (D): The optic disc was contoured, and the optic disc diameter was measured. (F): The type of the vessels was categorized into arteries or veins. (G–J): Multiple biomarkers on the extracted vasculature, including the vessel curvature, bifurcation feature, width, and the fractal for arteries and veins, were extracted, respectively. We measured the maximum, mean, median and minimum value for each biomarker, which yielded, in total, 96 biomarkers for each fundus image. At last, we built machine learning models to predict the corresponding CAD-RADS score.

• 

  Figure 2. The image quality of fundoscopy was graded by two ophthalmologists. The fundus images with sub-optimal quality were excluded in this study. (a) Satisfactory image; (b) sub-optimal image.

• 

  Figure 3. The retinal vasculature was manually annotated by using the “NeuronJ” plugin in ImageJ software. The vessel centerlines were traced on the vessel soft segmentation map. The parent-child relationship of the vessels was determined using the labelling system in “NeuronJ”.

• 

  Figure 4. (a) A vessel centerline is initialized by the vessel skeletonization. (b) An active contour model is built based on the centerline. (c) The contour iteratively grows and fits the left and right boundaries of the vessel. (d) The contour is cropped to obtain the left and right boundary (e). (f) The vessel width is measured by finding the average distances from one side to the other.

• 

  Figure 5. (A–E) The population graph was constructed in which the nodes represent the fundus images characterized by 96 retinal vascular biomarkers, and the edges were the similarity score determined by the age and gender of the subjects. A GraphSAGE network was applied to the constructed graph to predict CAD-RADS score of the subjects.