Deep learning model for intravascular ultrasound image segmentation with temporal consistency

Abstract

This study was conducted to develop and validate a deep learning model for delineating intravascular ultrasound (IVUS) images of coronary arteries.Using a total of 1240 40-MHz IVUS pullbacks with 191,407 frames, the model for lumen and external elastic membrane (EEM) segmentation was developed. Both frame- and vessel-level performances and clinical impact of the model on 3-year cardiovascular events were evaluated in the independent data sets. In the test set, the Dice similarity coefficients (DSC) were 0.966 +/- 0.025 and 0.982 +/- 0.017 for the lumen and EEM, respectively. Even at sites of extensive attenuation, the frame-level performance was excellent (DSCs > 0.96 for the lumen and EEM). The model (vs. the expert) showed a better temporal consistency for contouring the EEM. The agreement between the model- vs. the expert-derived cross-sectional and volumetric measurements was excellent in the independent retrospective cohort (all, intra-class coefficients > 0.94). The model-derived percent atheroma volume > 52.5% (area under curve 0.70, sensitivity 71% and specificity 67%) and plaque burden at the minimal lumen area site (area under curve 0.72, sensitivity 72% and specificity 66%) best predicted 3-year cardiac death and nonculprit-related target vessel revascularization, respectively. In the stented segment, the DSCs > 0.96 for contouring lumen and EEM were achieved. Applied to the 60-MHz IVUS images, the DSCs were > 0.97. In the external cohort with 45-MHz IVUS, the DSCs were > 0.96. The deep learning model accurately delineated vascular geometry, which may be cost-saving and support clinical decision-making.