Study population. Between January 2014 and July 2016, pre-procedural IVUS was conducted in 1657 stable and unstable angina patients with an angiographic diameter stenosis >40% on visual estimation at Asan Medical Center, Seoul, Republic of Korea. In patients who had IVUS pullbacks of two or more vessels, the vessel with the most severe stenosis on angiography was chosen. After excluding 417 cases with a stented lesion, chronic total occlusion, 20-MHz IVUS images, or technical errors in the image files, a final cohort of 1240 native coronary arteries in 1240 patients was enrolled for the development of the deep leaning model. The patients were randomly assigned to the training, validation and test sets at a ratio of 8:1:1. Thus, the current analysis included 987 IVUS pullbacks (152,448 frames) in the training set, 126 IVUS pullbacks (19,621 frames) in the validation set, and 127 IVUS pullbacks (19,338 frames) in the test set.
The protocol for this retrospective data analysis was approved by the institutional review board of Asan Medical Center and the requirement of written informed consent from the participants was waived.
Acquisition of IVUS. After intracoronary administration of 0.2 mg nitroglycerin, grayscale IVUS imaging was performed using a motorized transducer pullback (0.5 mm/s) and a commercial scanner (Boston Scientific/SCIMED, Minneapolis, MN) consisting of a rotating 40-MHz transducer within a 3.2-F imaging sheath. A region of interest was defined as the segment from the ostium to a point located 5-mm distal to the lesion (maximal plaque thickness ≥ 0.5 mm).
Model development. The lumen and vessel boundaries were labeled manually by experienced users in every IVUS frame with a 0.2-mm interval. Lumen segmentation was undertaken based on the interface between the lumen and the leading edge of the intima. A discrete interface at the border between the media and the adventitia corresponded approximately to the location of the EEM.
The overall workflow of the model development is shown in Figure 1. The adjacent 0.4-mm segments containing 13 frames were utilized for contouring a given target section. With the extraction of features from those frames, ResNet-50 generated a feature map with dimensions of 16x16x13. Transformer aggregated comprehensive information based on the similarities of the features across 13 frames, which enabled the model to attenuated frame-to-frame variabilities. The features were converted into polar coordinates, and were subsequently transformed into a segmentation mask (Supplemental Figure 1). The implementation details and data augmentation techniques are described in the Supplementary Appendix.
Each cross-sectional image was segmented into three compartments: (1) the adventitia, including the pixels outside the EEM (coded as “0”); (2) the lumen, including the pixels within the lumen border (coded as “1”); and (3) the plaque, including the pixels between the lumen border and the EEM (coded as “2”). To calibrate the pixel dimensions, grid lines were automatically applied in the IVUS images, and the pixel spacing was calculated for extracting the IVUS parameters.
To assess the model performance, the extent of overlap between the model-derived vs. the expert-measured lumen and EEM areas was assessed by three evaluation metrics (described in the Supplementary Appendix) including the Dice similarity coefficient (DSC), Jaccard index (JI) and Surface Dice similarity coefficient (SDSC). To exclude the potential clustering effect of multiple frames per vessel, the mean performance metrics calculated in each vessel were averaged in the test set.
Model validation. The vessel-level performances was retrospectively evaluated in the independent cohort of the Statin and Atheroma Vulnerability Evaluation trial (Supplementary appendix). Using computerized planimetry (EchoPlaque 3.0, Indec Systems, Mountain View, CA), quantitative IVUS analysis was conducted in accordance with the standards of the American College of Cardiology and the European Society of Cardiology.13 Using 111 pre-procedural IVUS pullbacks, the intra- and inter-observer variances in the core laboratory analysis were assessed by Expert 1 and 2.
In the lesions with extensive calcification or tissue attenuation, the frame-level performance was evaluated. Of 19,338 frames in the test set, 3,442 (17.8%) showed an arc of IVUS attenuation > 90˚ without an ultrasound signal behind a lipid-rich plaque or calcification.14-16 To evaluate the performance at bifurcation sites, 206 segments within the polygon of confluence (POC), a zone from the carina to the distal end of the proximal main branch, were also identified in the test set.17 The extent of overlap between the model-derived vs. the expert-measured lumen and EEM areas was assessed at the frame-level.
The model derived from native coronary arteries was applied to the stented segments in 165 vessels (132 for training, 16 for validation and 17 for testing) that were treated by stent implantation in Asan Medical Center, Seoul between July 2022 and December 2022. On the immediate post-stenting IVUS images, both the stent and EEM borders were manually labeled (Medilabel, Ingradient Inc., Seoul, Korea). The frame-level performance within the stented segment was evaluated before and after fine-tuning the model.
In addition, the model was applied to 60-MHz pre-procedural IVUS images (OptiCross HD, Boston Scientific Corporation, Marlborough, Massachusetts, USA) that were obtained at Asan Medical Center, Seoul between July 2022 and December 2022. The images were extracted and manually labeled at 30-frame intervals. The frame-level performance for lumen and EEM segmentation in 50 IVUS pullbacks (including 3254 frames) was assessed.
Clinical Validation. Between April 2011 and December 2013, 790 patients underwent both 40-MHz IVUS and FFR measurements for at least one nonculprit (untreated) coronary artery with angiographic diameter stenosis > 40% at the Asan Medical Center, Seoul, Korea. With the exclusion criteria (Supplementary appendix), 652 patients were finally included in the retrospective validation. In patients with IVUS pullbacks of ≥2 nonculprits, the major epicardial coronary artery with the lowest FFR value was preferentially chosen as the target. The primary endpoint was cardiac death, and the secondary endpoints were nonfatal myocardial infarction and target vessel revascularization (TVR) at 3-year follow-up (Supplementary appendix).
External validation. In 65 patients undergoing PCI between April 2022 and July 2023, a total of 65 pre-stenting IVUS pullbacks obtained by a 45-MHz IVUS catheter (Refinity, Philips Volcano, San Diego, CA, USA) were collected from Chung-Ang University Hospital, Seoul, Republic of Korea. The images at extracted 30-frame intervals were manually labeled. In 65 IVUS pullbacks (including 1731 frames), the frame-level performance for lumen and EEM segmentation was assessed.
Statistical analysis. The statistical analyses used for evaluating the patient and lesion characteristics were performed using SPSS (version 10.0, SPSS Inc., Chicago, IL, USA). All values were expressed as means ± 1 standard deviation (continuous variables) or as counts and percentages (categorical variables). Continuous variables were compared using unpaired t-tests. A p value <0.05 was considered statistically significant. Intra-class correlation coefficient was used to assess the agreement between the expert-measured vs. the model-derived values. The intra-class correlation coefficient value between 0.75 and 1.0 was considered to be ‘excellent’. The comparison between the expert- measured and the model-derived parameters was shown by Bland-Altman plot. Time-to-event data were presented as Kaplan-Meier estimates and compared using the log-rank test at 3- and 5-year follow-ups. Survival curves were constructed using Kaplan-Meier estimates and compared by a Cox proportional hazard regression model.