MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Imaging General ePoster Discussion
  5. Initial Evaluation of the Use of a Convolutional Neural Network to Determine Coronary Artery Disease Severity Using Computed Tomography Angiography

Initial Evaluation of the Use of a Convolutional Neural Network to Determine Coronary Artery Disease Severity Using Computed Tomography Angiography

A Podgorsak1, 2*, K Sommer1, 2 , V Iyer3 , M Wilson1 , U Sharma1 , K Kumamaru3 , F Rybicki4 , D Mitsouras4 , E Angel5 , C Ionita1, 2 , (1) SUNY Buffalo, Buffalo, NY, (2) Canon Stroke and Vascular Research Center, Buffalo, NY, (3) Juntendo University, Tokyo, (4) University of Ottawa, Ottawa, ON, (5) Canon Medical Systems, Tustin, CA

Presentations

(Sunday, 7/14/2019) 4:30 PM - 5:00 PM

Room: Exhibit Hall | Forum 9

Purpose: To quantitatively assess the performance of a convolutional neural network (CNN) at classifying hemodynamic significance of coronary artery stenoses using curved multi-planar reconstructions (CMPR) of the coronary artery.

Methods: 50 coronary tree CT angiographies were collected at 70% of the R-R cardiac cycle. CMPRs were generated using different slice thicknesses (0.4 and 5.0 mm) considering four different rotational views around the vessel centerline per CTA for a total dataset size of 200 for each slice thickness. The dataset was split into a training cohort numbering 125 and a testing cohort numbering 75. Fractional flow reserve (FFR) values were measured invasively to create a labeled dataset. A CNN with custom architecture was developed in Keras to classify input CMPRs by the hemodynamic significance of the lesion. Additionally, the network synthesized class activation maps (CAMs) so that the most salient features in the CMPR data could be visualized. Network performance on the test cohort and the impact of different CMPR slice thickness was assessed using the area under receiver operating characteristic curves (AUC), classification accuracy, and a Student’s T-Test.

Results: Mean classification accuracy over the validation cohort was 59.9% (95% confidence interval, 54.2%-65.6%) and 68.1% (63.8%-72.4%) for the 0.4 and 5.0 mm CMPRs respectively. AUC was 0.619 (0.554-0.672) and 0.718 (0.675-0.773) for the 0.4 and 5.0 mm CMPRs respectively. There was a statistically significant advantage to using the 5.0 mm thick CMPR slices over the 0.4 mm CMPR slices for classifying coronary disease severity in terms of both classification accuracy (p = 0.004) and AUC (p = 0.01). CAM results indicated calcium burden and vessel geometry were the most salient features in the CMPR data.

Conclusion: This work indicates the potential clinical utility of a CNN to predict the hemodynamic significance of coronary artery stenoses using CMPRs of the coronary artery.

Funding Support, Disclosures, and Conflict of Interest: James H. Cummings Foundation, Canon Medical Systems

Keywords

CAD, Angiography, Computer Vision

Taxonomy

IM/TH- Image Analysis (Single modality or Multi-modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email

Additional Links