Abstract
Background: Although coronary CT angiography identifies a coronary stenosis, anatomical stenosis correlates moderately with the identification of ischemia to a specific vessel, which is the current reference standard and can only be determined by invasive fractional flow reserve. Phenotype of atherosclerotic plaque is another aortic parameter of ischemia that can be automatically quantified using deep learning. This was an assessment of a Deep-Learning Plaque Phenotyping (DLPP) model for predicting vessel specific ischemia. Methods: This is a retrospective diagnostic accuracy study, in which 787 vessels of 384 patients who underwent CTA and FFR were analysed. Phenotype of plaques was quantified using a deep learning model and the presence of vesselspecific ischemia (defined as fractional flow reserve at or below 0.80) was determined by using a classifier. DeLong test was used to compare discrimination vs stenosis alone, and calibration test performed. Results: 40% of vessels had ischemia. The deep-learning plaque model discriminated ischemia at an area under the curve of 0.86 (95% CI 0.81 to 0.91) with a sensitivity of 85%, a specificity of 77% and negative predictive value of 90%, significantly better than stenosis alone (at an area under the curve of 0.73, DeLong p<0.001). The strongest predictor was low-attenuation plaque volume and the model had a well calability (Hosmer-Lemeshow p = 0.49). Conclusion: Phenotyping plaques using deep-learning performed better than stenosis alone to predict vessel-specific ischemia and could therefore serve as a non-invasive gateway to invasive assessment.
Keywords: Coronary CT angiography, Deep learning, Plaque phenotyping, Fractional flow reserve, Vessel-specific ischemia, Low-attenuation plaque, Machine learning