Development and Validation of Deep Learning Model for Intravascular Ultrasound Image Segmentation

Development and Validation of Deep Learning Model for Intravascular Ultrasound Image Segmentation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and Validation of Deep Learning Model for Intravascular Ultrasound Image Segmentation Hyeonmin Kim, June-Goo Lee, Gyu-Jun Jeong, Geunyoung Lee, Hyunseok Min, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4633591/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 27 Aug, 2024 Read the published version in The International Journal of Cardiovascular Imaging → Version 1 posted 16 You are reading this latest preprint version Abstract Aims. This study was conducted to develop and validate a deep learning model for delineating intravascular ultrasound (IVUS) images of coronary arteries. Materials and Methods. Using atotal 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. Results. 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 good in the independent retrospective cohort (all, intra-class coefficients > 0.94). The model-derived PAV>52.5% (area under curve 0.70, sensitivity 71% and specificity 67%) and plaque burden at the MLA 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 by fine-tuning. 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. Conclusion. The deep learning model accurately delineated vascular geometry, which may be cost-saving and support clinical decision-making. intravascular ultrasound segmentation deep learning coronary artery disease Figures Figure 1 Figure 2 Figure 3 Introduction Intravascular ultrasound (IVUS) is a generalized intracoronary imaging modality that provides morphological information to the operators about the severity of stenosis, atheroma burden and plaque characteristics. IVUS is also useful for device sizing and stent optimization by correcting for suboptimal findings such as stent underexpansion, edge problems and procedure-related complications, which help prevent early and late stent failure. 1–3 Although the clinical impact of IVUS-guided percutaneous coronary intervention (PCI) has been well established, 4,5 IVUS still remains underused. Interpretation and measurement of IVUS images is time-consuming and demands adequately trained personnel, which partly explains the low penetration rate of IVUS in the real world. Precise delineation of the lumen and external elastic membrane (EEM) borders is essential for a morphological assessment. However, frame-by-frame manual annotation is an error-prone task with high inter- and intra-observer variabilities. Especially in the presence of blood speckling obscuring the luminal border, tissue attenuation by intra-plaque calcium or lipid, or various image artifacts, it is tricky to extract the contour without consideration of the morphologic characteristics of the adjacent segments. Even with the selection of every 60th (1-mm interval) or 30th (0.5-mm interval) frame, volumetric analysis in an entire IVUS pullback containing thousands of images usually takes several hours. When derived from one selected image, the traditional IVUS criteria such as minimal lumen or stent area and plaque burden may be insufficient to represent the status of a whole vascular segment. Therefore, a program for automatic segmentation of the entire sequence of pullback images is necessary to facilitate the on-site utilization of IVUS and to stratify future cardiovascular risks. Although a variety of methodologies for semantic segmentation have been applied, programs for automated IVUS analysis have not been widely employed in real practice, which is possibly due to their suboptimal performance, relatively small number of training samples, and incomplete clinical validation. 6–11 It is uncertain if the models based on the 40-MHz IVUS images of native coronary arteries can be applied to 60-MHz IVUS images or stented lesions. Moreover, a lack of external validation limits the generalization of the developed programs. Thus, the aims of the current study were to 1) develop a 40-MHz IVUS-based deep learning model for delineating the lumen and EEM borders of native coronary arteries, 2) validate the performance and clinical impact of the model, and 3) evaluate whether the model can be applied to 60-MHz IVUS images or stented segments. Materials and methods 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. Results Clinical and lesion characteristics. In the study cohort for model development, the mean age was 64.3 ± 9.5 years, and 75% were men. The target vessels were the left anterior descending artery in 73%, the left circumflex artery in 5%, the right coronary artery in 18%, the ramus intermedius in 2%, and the left main coronary artery in 2%. Frame-level performance. Based on the evaluation metrics, the frame-level performance for lumen and EEM segmentation are summarized in Table 1. The model performance was also shown with the inclusion of 3,442 (17.8%) frames that showed an arc of IVUS signal loss > 90˚ behind attenuated or calcified plaque. Figure 2 compares the frame-to-frame variabilities of the model- vs. the expert-derived segmentation. To exclude the potential clustering effect of multiple frames in a vessel, the mean performance indices calculated in each of the 127 IVUS pullbacks in the test set were averaged. The averages of the mean DSC were 0.967 ± 0.007 and 0.982 ± 0.006 for the lumen and EEM segmentation, respectively. The averages of the mean JI were 0.938 ± 0.013 and 0.966 ± 0.011, respectively. In addition, the averages of the mean SDSC were 0.869 ± 0.055 and 0.910 ± 0.011, respectively. The inference time was 0.029 second per frame, and the model-derived segmentation for 50-mm length (including 3000 frames) required 88 seconds. In comparison, manual segmentation with 0.2-mm interval took 187.5 minutes on average. Vessel-level performance. The vessel-level performance of the model was evaluated in the cohort of the previous trial. The target vessels were the left anterior descending artery in 39%, the left circumflex artery in 19%, the right coronary artery in 40%, and the ramus intermedius in 2%. The lesion length was 25.9±8.7 mm, and plaque burden at the minimal lumen area (MLA) site was 67.6±10.9. There were no significant differences in the IVUS parameters measured by the model vs. the expert (Table 2). Overall the agreement between the model-derived vs. the expert-measured parameters was excellent (Figure 3). However, the mean MLA derived by the model (vs. measured by the expert) was significantly smaller (4.1±1.8 mm 2 vs. 4.3±2.0 mm 2 , p<0.001). The model-derived MLA was smaller than that measured by the expert in 78.8% of the cases. The intra- and inter-observer variations in the core laboratory analysis are shown in Supplemental Table 1, Supplemental Figure 2 and Supplemental Figure 3. Clinical Validation. In the retrospective cohort including 652 patients, baseline characteristics of patients and lesions are summarized in Supplemental Table 2. The median follow-up was 47.8 months (IQR 40.1–67.8 months). At 3 years, cardiac death and acute MI occurred in 21 (3.2%) and 5 (0.8%) patients, respectively. Noncuprit-related TVR was performed in 28 (4.3%) patients during 3-year follow-up. Table 3 compares the model-derived IVUS measurements between patients with vs. without 3-year cardiac death and TVR. There were no significant differences in the IVUS measurements or FFR between patients with vs. without acute MI. Among the IVUS variables, PAV>52.5% (area under curve 0.70, sensitivity 71% and specificity 67%) and plaque burden at the MLA site (area under curve 0.72, sensitivity 72% and specificity 66%) best predicted 3-year cardiac death and TVR, respectively (Supplemental Table 3). On the Kaplan-Meier curves (Supplemental Figures 4), the cardiac death-free survival rate at 3 years was significantly lower with PAV > 52.5% vs. ≤ 52.5% (93.4% vs. 98.6%, log-rank p 76.5% vs. ≤ 76.5% (91.5% vs. 98.1%, log-rank p <0.001). Stented segment. Table 4 shows the frame-level performances in the stented segments before and after fine-tuning of the model. For lumen segmentation, the DSC was increased from 0.918 ± 0.025 pre-tuning to 0.975 ± 0.010 post-tuning in the test set. To contour the EEM border, the DSC was improved from 0.928 ± 0.059 to 0.963 ± 0.026 by fine-tuning the model. 60-MHz IVUS images. When the model was tested on the 60-MHz IVUS images, the DSCs were 0.978 ± 0.024 and 0.985 ± 0.017 for the lumen and EEM, respectively. The JIs were 0.958 ± 0.040 and 0.972 ± 0.030, respectively. The SDSCs were 0.970 ± 0.086 and 0.967 ± 0.098, respectively. External validation. The model was also tested in the external data. For lumen and EEM segmentation, the DSCs were 0.960 ± 0.049 and 0.979 ± 0.021, respectively. The JIs were 0.927 ± 0.078 and 0.960 ± 0.038, respectively. The SDSCs were 0.875 ± 0.181 and 0.892 ± 0.190, respectively. Discussion Although a variety of methodologies for automatic IVUS segmentation have been applied, there is a lack of models that are widely used in real practice. Their suboptimal performance can be explained by several reasons. With the requirement of a large data set for supervised deep learning, an insufficient number of manually-labeled training samples is one of the reasons for their poor accuracy. Moreover, there are fundamental pitfalls of IVUS images, including their limited spatial resolution, to-and-fro motion of the imaging catheter and various artifacts that make it challenging to delineate the vascular geometry. Especially at sites of tissue attenuation behind calcium or lipid, it has been considered that there is a need to utilize multiple adjacent frames for contouring the presumptive vessel border. However, it remains uncertain how to explicitly exploit a complex series of cross-sectional images. As a data-driven approach, convolutional neural networks that have been designed to automatically and adaptively ascertain the spatial hierarchies of features and have been adopted in many computer vision applications. 6-11 For IVUS segmentation, Yang et al. previously demonstrated that Dual Path U-Net had superior performance (JI 0.823 and 0.775 for lumen and EEM segmentation, respectively) relative to conventional computer vision-based approaches. 9 Nishi et al also developed DeepLabv3 with a modification of the encoder component. 10 Due to their exclusion of images containing tissue attenuation ≥90° of arc, image artifacts and large side branches, their model could not be applied to general use. By tracing the 4 longitudinal cutting planes spaced by 45 degrees (as the ground truths), Ziemer et al. developed a model for IVUS segmentation at the dataset level. 11 In the presence of a saw-tooth artifact throughout the longitudinal view, electrocardiography-gated selection of the end-diastolic frames is mandatory, which raises concerns about inadequate temporal resolution. In addition, electrocardiogram-synchronized images are currently available only for 20-MHz IVUS system. In this study, we proposed a deep learning model combining ResNet-50 with a Transformer architecture which is adept at capturing long-range dependencies. This combination facilitates the interaction of extracted features across both spatial and temporal dimensions. Additionally, the implementation of a polar coordinate transformation aids in maintaining the integrity of the segmentation masks. With a dataset comprising 191,407 frames that were manually labeled, our model significantly advanced both frame- and vessel-level performance. Even at sites of extensive calcification or tissue attenuation, our model showed excellent performance (DSC > 0.96) for contouring both the lumen and the EEM. Frame-by-frame manual delineation of the invisible EEM with a consistent criterion is challenging, which might lead to high frame-to-frame variability. However, this current model achieved a much better temporal consistency of EEM segmentation by incorporating the overall geometrical information from the series of adjacent frames. Although the precise measurement of MLA is important to assess lesion severity, determine device size, and predict future clinical outcomes, 17,18 there is large variability in the selection of the MLA frame by visually inspecting the pullback. In our study, the model- (vs. the expert-) derived MLA was smaller in most cases. The model-based whole-frame analysis enables us to select the right frame of the MLA. By eliminating the effect of human subjectivity or uncertainty of interpretation, the automatic analysis may be useful for the accurate diagnosis. Our study validated clinical impact of the model-derived measurements. Among the IVUS parameters, PAV >52.5% and plaque burden at the MLA site >76.5% best predicted 3-year cardiac death and nonculprit-related TVR, respectively. The occurrence of cardiac death may depend on the overall atheroma burden in the entire vascular segment, while TVR is more likely determined by the localized disease status. Because metallic strut shadows usually interfere with the full visualization of the EEM, the application of the model derived from native coronary arteries to stented vessels caused the performance degradation. Nonetheless, fine-tuning by using a data set of stent images improved the performance for contouring both stent (DSC 0.975 ± 0.010) and the EEM (DSC 0.963 ± 0.026). Moreover, the model based on the 40-MHz IVUS images consistently showed high accuracies for contouring the 60-MHz images. External validation of a deep learning model is an essential step to determine its reproducibility and generalizability to different devices and settings, and to patients with diverse lesion characteristics. Applied to the 45-MHz IVUS images in the external PCI cohort, good performance of our model was demonstrated. By saving time and expenses in core laboratories, the program potentially facilitates the on-site utilization of IVUS and real-time decision-making during PCI. Study limitations. First, the current model needs to be validated in all of the commercially available systems, and tested on images obtained by different pullback speed or manual pullback. Second, even with its good performance for stent segmentation, this model cannot be used for assessing in-stent-restenosis. For a complete post-stenting evaluation, further tuning of the model should be performed for the meticulous delineation of neointima, tissue prolapse, stent malapposition and intimal dissection. In addition, model training in lesion subsets with intimal disruption, dissection, intraluminal thrombus and nodular calcification is necessary. With a lack of expert consensus for the segmentation of the complex bifurcation geometry, it was challenging to perform data labeling with incoherent criteria, which might be responsible for the suboptimal performance within the POC of bifurcation. Moreover, further study to evaluate possible ethnic differences in the model’s performance is required. Even though there is a concern about the reproducibility of measurement potentially affecting the quality of data labeling, the intra- and inter-observer variances in our core laboratory analysis were not considerable. Conclusion This deep learning-based model precisely delineated vascular geometry and improved temporal consistency. With a good performance, it can be applied to the stented segment and 60-MHz IVUS images. The model-derived IVUS measurements based on the whole-frame analysis had prognostic implication for the occurrence of cardiac mortality and TVR at long term. Declarations Informed consent and patient details. The authors declare that this report does not contain any personal information that could lead to the identification of the patients. Author contributions. All authors attest that they meet the current International Committee of Medical Journal Editors (ICMJE) criteria for Authorship. Hyeonmin Kim: model development, data analysis, writing and edit the manuscript. June-Goo Lee and Gyu-Jun Jeong: methodology, supervision and review Geunyoung Lee, H Cho and H Min: development of model, data processing, software and methodology Daegyu Min: methodology, review and edit the manuscript. SW Lee, Jun Hwan Cho, and Sungsoo Cho: validation, review and edit the manuscript. SJ Kang: Conceptualization and design of study, data curation, methodology, supervision, validation, Funding, development of model, and writing and edit the original draft. Declaration of Competing interest Kim H & Lee G is an employee of Mediwhale Inc., Seoul, Korea. Min D is an employee of and Ingradient Inc., Seoul, Korea. Other authors report no conflicts of interest regarding this manuscript. Funding support and author disclosures This study was supported by grants from the Ministry of Science and ICT (NRF-2021R1A2C2006831) and the Asan Institute for Life Sciences, Asan Medical Center, Seoul, Republic of Korea (2021IP0071-1). Acknowledgments We thank Mediwhale Inc. and Ingradient Inc. for their technical support. Data availability statemen. The datasets generated and/or analyzed during the current study are not publicly available because permission of sharing patient data was not granted by the Institutional Review Board but are available from the corresponding author on reasonable request. References Fujii K, Carlier SG, Mintz GS, Yang YM, Moussa I, Weisz G, Dangas G, Mehran R, Lansky AJ, Kreps EM, Collins M, Stone GW, Moses JW, Leon MB (2005) Stent underexpansion and residual reference segment stenosis are related to stent thrombosis after sirolimus-eluting stent implantation: an intravascular ultrasound study. J Am Coll Cardiol 45(7):995-998. Okabe T, Mintz GS, Buch AN, Roy P, Hong YJ, Smith KA, Torguson R, Gevorkian N, Xue Z, Satler LF, Kent KM, Pichard AD, Weissman NJ, Waksman R (2007) Intravascular ultrasound parameters associated with stent thrombosis after drug-eluting stent deployment. Am J Cardiol 100(4):615-620. Liu X, Doi H, Maehara A, Mintz GS, Costa Jde R Jr, Sano K, Weisz G, Dangas GD, Lansky AJ, Kreps EM, Collins M, Fahy M, Stone GW, Moses JW, Leon MB, Mehran R. A (2009) A volumetric intravascular ultrasound comparison of early drug-eluting stent thrombosis versus restenosis. JACC Cardiovasc Interv 2(5): 428-434. Lee JM, Choi KH, Song YB, Lee JY, Lee SJ, Lee SY, Kim SM, Yun KH, Cho JY, Kim CJ, Ahn HS, Nam CW, Yoon HJ, Park YH, Lee WS, Jeong JO, Song PS, Doh JH, Jo SH, Yoon CH, Kang MG, Koh JS, Lee KY, Lim YH, Cho YH, Cho JM, Jang WJ, Chun KJ, Hong D, Park TK, Yang JH, Choi SH, Gwon HC, Hahn JY; RENOVATE-COMPLEX-PCI Investigators (2023) Intravascular Imaging-Guided or Angiography-Guided Complex PCI. N Engl J Med 388(18):1668-1679. Witzenbichler B, Maehara A, Weisz G, Neumann FJ, Rinaldi MJ, Metzger DC, Henry TD, Cox DA, Duffy PL, Brodie BR, Stuckey TD, Mazzaferri EL Jr, Xu K, Parise H, Mehran R, Mintz GS, Stone GW (2014) Relationship between intravascular ultrasound guidance and clinical outcomes after drug-eluting stents: the assessment of dual antiplatelet therapy with drug-eluting stents (ADAPT-DES) study. Circulation 129(4):463-470. Bajaj R, Huang X, Kilic Y, Ramasamy A, Jain A, Ozkor M, Tufaro V, Safi H, Erdogan E, Serruys PW, Moon J, Pugliese F, Mathur A, Torii R, Baumbach A, Dijkstra J, Zhang Q, Bourantas CV (2021) Advanced deep learning methodology for accurate, real-time segmentation of high-resolution intravascular ultrasound images. Int J Cardiol 339:185-191. Zhu F, Gao Z, Zhao C, Zhu H, Nan J, Tian Y, Dong Y, Jiang J, Feng X, Dai N, Zhou W (2022) A Deep Learning-based Method to Extract Lumen and Media-Adventitia in Intravascular Ultrasound Images. Ultrason Imaging 44(5-6):191-203. Mendizabal-Ruiz EG, Rivera M, Kakadiaris IA (2013) Segmentation of the luminal border in intravascular ultrasound B-mode images using a probabilistic approach. Med Image Anal 17(6):649-670. Yang J, Faraji M, Basu A (2019) Robust segmentation of arterial walls in intravascular ultrasound images using Dual Path U-Net. Ultrasonics 96:24-33. Nishi T, Yamashita R, Imura S, Tateishi K, Kitahara H, Kobayashi Y, Yock PG, Fitzgerald PJ, Honda Y (2021) Deep learning-based intravascular ultrasound segmentation for the assessment of coronary artery disease. Int J Cardiol 333:55-59. Ziemer PGP, Bulant CA, Orlando JI, Maso Talou GD, Álvarez LAM, Guedes Bezerra C, Lemos PA, García-García HM, Blanco PJ (2020) Automated lumen segmentation using multi-frame convolutional neural networks in intravascular ultrasound datasets. Eur Heart J Digit Health 1(1):75-82. Park SJ, Kang SJ, Ahn JM, Chang M, Yun SC, Roh JH, Lee PH, Park HW, Yoon SH, Park DW, Lee SW, Kim YH, Lee CW, Mintz GS, Han KH, Park SW (2016) Effect of Statin Treatment on Modifying Plaque Composition: A Double-Blind, Randomized Study. J Am Coll Cardiol 67(15):1772-1783. Mintz GS, Nissen SE, Anderson WD, Bailey SR, Erbel R, Fitzgerald PJ, Pinto FJ, Rosenfield K, Siegel RJ, Tuzcu EM, Yock PG (2001) American College of Cardiology Clinical Expert Consensus Document on Standards for Acquisition, Measurement and Reporting of Intravascular Ultrasound Studies (IVUS): a report of the American College of Cardiology Task Force on Clinical Expert Consensus Documents. J Am Coll Cardiol 37(5):1478-1492. Kang SJ, Mintz GS, Pu J, Sum ST, Madden SP, Burke AP, Xu K, Goldstein JA, Stone GW, Muller JE, Virmani R, Maehara A (2015) Combined IVUS and NIRS detection of fibroatheromas: histopathological validation in human coronary arteries. JACC Cardiovasc Imaging 8(2):184-194. Pu J, Mintz GS, Biro S, Lee JB, Sum ST, Madden SP, Burke AP, Zhang P, He B, Goldstein JA, Stone GW, Muller JE, Virmani R, Maehara A (2014) Insights into echo-attenuated plaques, echolucent plaques, and plaques with spotty calcification: novel findings from comparisons among intravascular ultrasound, near-infrared spectroscopy, and pathological histology in 2,294 human coronary artery segments. J Am Coll Cardiol 63(21):2220-2233. Kang SJ, Ahn JM, Han S, Park DW, Lee SW, Kim YH, Lee CW, Park SW, Mintz GS, Park SJ (2016) Multimodality imaging of attenuated plaque using grayscale and virtual histology intravascular ultrasound and optical coherent tomography. Catheter Cardiovasc Interv 88(1):E1-11. Kang SJ, Mintz GS, Oh JH, Park DW, Lee SW, Kim YH, Lee CW, Park SW, Park SJ (2013) Intravascular ultrasound assessment of distal left main bifurcation disease: the importance of the polygon of confluence of the left main, left anterior descending, and left circumflex arteries. Catheter Cardiovasc Interv 82(5):737-745. Kang SJ, Lee JY, Ahn JM, Mintz GS, Kim WJ, Park DW, Yun SC, Lee SW, Kim YH, Lee CW, Park SW, Park SJ (2011) Validation of intravascular ultrasound-derived parameters with fractional flow reserve for assessment of coronary stenosis severity. Circ Cardiovasc Interv 4(1):65-71. Stone GW, Maehara A, Lansky AJ, de Bruyne B, Cristea E, Mintz GS, Mehran R, McPherson J, Farhat N, Marso SP, Parise H, Templin B, White R, Zhang Z, Serruys PW; PROSPECT Investigators (2011) A prospective natural-history study of coronary atherosclerosis. N Engl J Med 364(3);226-235. Tables Table 1. Frame-level performance for lumen and EEM segmentation For lumen segmentation* For EEM segmentation* DSC JI SDSC DSC JI SDSC Overall Validation set 0.966 ± 0.025 0.936 ± 0.042 0.871 ± 0.142 0.982 ± 0.017 0.965 ± 0.029 0.911 ± 0.134 Test set 0.968 ± 0.025 0.938 ± 0.042 0.874 ± 0.145 0.982 ± 0.017 0.966 ± 0.031 0.910 ± 0.141 Segments with extensive IVUS attenuation † 0.968 ± 0.025 0.938 ± 0.041 0.889 ± 0.136 0.973 ± 0.023 0.950 ± 0.041 0.803 ± 0.191 Within the POC ‡ 0.941 ± 0.039 0.892 ± 0.067 0.747 ± 0.147 0.968 ± 0.026 0.940 ± 0.046 0.808 ± 0.125 EEM = external elastic membrane, DSC = Dice similarity coefficient, JI = Jaccard index, SDSC = Surface Dice similarity coefficient * Including all frames at 0.2-mm (12 frames) intervals † Including 123 segments with 3442 frames (in the test set) that showed an arc of IVUS signal loss > 90˚ behind the attenuated or calcified plaque ‡ 206 bifurcation segments defined as the polygon of confluence between the carina and the distal end of the proximal main branch Table 2 . Comparison of model- vs. expert-derived measurements. Expert Model Cross-sectional Lesion length 25.9±8.7 25.8±8.7 Minimal lumen area, mm 2 4.3±2.0 4.1±1.8* EEM at the MLA site, mm 2 13.6±4.7 13.3±4.3 P+M at the MLA site, mm 2 9.3±3.9 9.2±3.5 Plaque burden at the MLA site, % 67.6±10.9 68.6±10.2 Mean EEM diameter, mm 4.1±0.7 4.1±0.7 Volumetric Lumen volume, mm 3 194.0±92.7 195.3±92.3 EEM volume, mm 3 401.6±186.2 402.8±183.2 Plaque volume, mm 3 207.6±107.3 207.5±103.1 Percent atheroma volume (%) 51.3±8.3 51.2±7.9 *p values <0.05 (vs. expert) Table 3. Model-derived IVUS parameters and FFR between patients with and without 3-year clinical events. Cardiac death Target vessel revascularization (+) (-) (+) (-) IVUS measurements MLA, mm 2 2.7 (2.0 – 4.3) 2.9 (2.2 – 3.7) 2.3 (2.0 – 3.0 ) 2.9 (2.2 – 3.7) # EEM area at the MLA site, mm 2 9.6 (7.1 – 14.9) 10.9 (8.0 – 14.4) 11.6 (9.9 – 15.9) 10.9 (7.8 – 14.3) P+M area at the MLA site , mm 2 7.1 (5.1 – 10.8) 8.0 (5.2 – 10.8) 9.4 (7.5 – 12.8) 7.8 (5.1 – 10.7) # Plaque burden at the MLA site, % 75.9 (69.1 – 77.6) 73.4 (65.8 – 78.8) 80.1 (72.1 – 83.3) 73.3 (65.8 – 78.4) # Lumen volume, mm 3 269.1 (220.9 – 489.4) 310.6 (208.8 – 416.5) 247.0 (153.1 – 350.8) 312.2 (213.4 – 426.4) # EEM volume, mm 3 614.4 (474.9 – 1106.8) 609.1 (421.4 – 819.3) 518.4 (365.3 – 725.7) 612.7 (433.9 – 826.4) P+M volume, mm 3 382.4 (249.6 – 587.0) 290.8 (206.9 – 401.7)* 232.5 (199.1 – 380.1) 296.2 (208.7 – 407.8) Percent atheroma volume, % 55.7 (49.7 – 58.7) 48.8 (43.5 – 54.2)* 54.8 (45.0 – 59.6) 48.8 (43.5 – 54.2) # Length of ROI, mm 53.2 (37.6 – 65.9) 45.7 (34.3 – 56.1) 41.4 (27.1 – 51.6) 45.9 (34.6 – 56.3) FFR at maximal hyperemia 0.86 (0.83 – 0.89) 0.86 (0.81 – 0.90) 0.81 (0.74 – 0.84) 0.86 (0.82 – 0.90) # MLA= minimal lumen area, EEM=external elastic membrane, P+M= plaque + media, ROI= region of interest, * p value < 0.05 vs. 3-year cardiac death (+) group (by Mann-Whitney), # p value < 0.05 vs. 3-year target vessel revascularization (+) group (by Mann-Whitney) Table 4. Frame-level performance in stented segments For lumen segmentation* For EEM segmentation* DSC JI SDSC DSC JI SDSC Pre-tuning Validation set 0.931±0.023 0.872±0.039 0.371±0.208 0.949±0.051 0.908±0.084 0.621±0.292 Test set 0.918±0.025 0.850±0.042 0.281±0.179 0.928±0.059 0.870±0.096 0.458±0.307 Post-tuning Validation set 0.976±0.010 0.953±0.018 0.875±0.116 0.971±0.026 0.945±0.047 0.758±0.239 Test set 0.975±0.010 0.952±0.018 0.881±0.110 0.963±0.026 0.930±0.046 0.680±0.239 *Including all frames at 0.2-mm (12 frames) intervals EEM = external elastic membrane, DSC = Dice similarity coefficient, JI = Jaccard index, SDSC = Surface Dice similarity coefficient Additional Declarations Competing interest reported. Kim H & Lee G is an employee of Mediwhale Inc., Seoul, Korea. Min D is an employee of and Ingradient Inc., Seoul, Korea. Other authors report no conflicts of interest regarding this manuscript. Supplementary Files segmentationsupplIJCImaging.doc Supplemental Figure 1. Architecture of ResNet-50 & Transformer. Supplemental Figure 2. Bland-Altman plots between inter-observers of IVUS measurements. Supplemental Figure 3. Correlations between inter-observer variabilities of IVUS measurements. Supplemental Figure 4. Kaplan-Meier curves for event-free survival. Cite Share Download PDF Status: Published Journal Publication published 27 Aug, 2024 Read the published version in The International Journal of Cardiovascular Imaging → Version 1 posted Editorial decision: Revision requested 12 Jul, 2024 Reviews received at journal 12 Jul, 2024 Reviews received at journal 11 Jul, 2024 Reviews received at journal 11 Jul, 2024 Reviews received at journal 10 Jul, 2024 Reviews received at journal 05 Jul, 2024 Reviewers agreed at journal 01 Jul, 2024 Reviewers agreed at journal 28 Jun, 2024 Reviewers agreed at journal 28 Jun, 2024 Reviewers agreed at journal 28 Jun, 2024 Reviewers agreed at journal 28 Jun, 2024 Reviewers agreed at journal 28 Jun, 2024 Reviewers invited by journal 28 Jun, 2024 Editor assigned by journal 25 Jun, 2024 Submission checks completed at journal 25 Jun, 2024 First submitted to journal 25 Jun, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4633591","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":326238490,"identity":"6e398529-bb51-4a67-9fef-ae0771f1effa","order_by":0,"name":"Hyeonmin Kim","email":"","orcid":"","institution":"Pohang University of Science and Technology (POSTECH)","correspondingAuthor":false,"prefix":"","firstName":"Hyeonmin","middleName":"","lastName":"Kim","suffix":""},{"id":326238492,"identity":"9099cf90-aa3d-4acb-888e-1df1372b7f15","order_by":1,"name":"June-Goo Lee","email":"","orcid":"","institution":"Asan Institute for Life Sciences","correspondingAuthor":false,"prefix":"","firstName":"June-Goo","middleName":"","lastName":"Lee","suffix":""},{"id":326238494,"identity":"51519d9c-5403-4688-9305-3e21516f3408","order_by":2,"name":"Gyu-Jun Jeong","email":"","orcid":"","institution":"Asan Institute for Life Sciences","correspondingAuthor":false,"prefix":"","firstName":"Gyu-Jun","middleName":"","lastName":"Jeong","suffix":""},{"id":326238495,"identity":"bb0c764c-bc30-431e-b0aa-94af7f6e8e59","order_by":3,"name":"Geunyoung Lee","email":"","orcid":"","institution":"Mediwhale Inc","correspondingAuthor":false,"prefix":"","firstName":"Geunyoung","middleName":"","lastName":"Lee","suffix":""},{"id":326238497,"identity":"05d3c430-49de-40f1-b1ab-6f96430b88ee","order_by":4,"name":"Hyunseok Min","email":"","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hyunseok","middleName":"","lastName":"Min","suffix":""},{"id":326238499,"identity":"a4f1b6a7-5fbb-49d0-ad9f-564f766abd78","order_by":5,"name":"Hyungjoo Cho","email":"","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hyungjoo","middleName":"","lastName":"Cho","suffix":""},{"id":326238500,"identity":"f96e2429-835a-4f6b-ac83-28f6bbf0afa2","order_by":6,"name":"Daegyu Min","email":"","orcid":"","institution":"Ingradient Inc","correspondingAuthor":false,"prefix":"","firstName":"Daegyu","middleName":"","lastName":"Min","suffix":""},{"id":326238501,"identity":"ed655493-0523-45ee-a646-ea95287c08aa","order_by":7,"name":"Seung-Whan Lee","email":"","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Seung-Whan","middleName":"","lastName":"Lee","suffix":""},{"id":326238502,"identity":"bd5baa51-d836-43ad-8f37-eee2dc1408b6","order_by":8,"name":"Jun Hwan Cho","email":"","orcid":"","institution":"Chung-Ang University Gwangmyeong Hospital, Chung-Ang University College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Jun","middleName":"Hwan","lastName":"Cho","suffix":""},{"id":326238503,"identity":"236a500b-76ca-49b8-8980-db390cc132e0","order_by":9,"name":"Sungsoo Cho","email":"","orcid":"","institution":"Yonsei University College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Sungsoo","middleName":"","lastName":"Cho","suffix":""},{"id":326238504,"identity":"234aaec3-dd45-49fa-8773-0e207a0b5c0b","order_by":10,"name":"Soo-Jin Kang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIie3PvwqCQBzA8Z8IN93hmhj5BEHR2L9XSQRbGx2NwBZprhdxPjg4l6vWwKVHEFykDDodGtW2oPsOP7jjPvA7AJXqJ+PVmH2OfcDdiPc54g4EVYN9Qewtd/LN42oP91TPMp9Jcr43khHlzDod0nEsVsg8CknIftRMIAksEqVafAOkk1ASA7UsFiS7J4kuS0n0/NWFAOXcwgV1JAFLqwgJm4X8izclgevGwgnNSKwxwrxlsSOfpLhczOOEsazwpwMDe80EehRAq1fRgvqi7ScARvWwbH2mUqlU/9wbfzNGq91OwmcAAAAASUVORK5CYII=","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Soo-Jin","middleName":"","lastName":"Kang","suffix":""}],"badges":[],"createdAt":"2024-06-25 05:06:58","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4633591/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4633591/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10554-024-03221-9","type":"published","date":"2024-08-27T15:56:56+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":60811112,"identity":"45458f52-998f-441b-9981-fff2543c91cd","added_by":"auto","created_at":"2024-07-22 10:53:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":606829,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDevelopment of a deep learning model for IVUS segmentation.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4633591/v1/03166a04cdf88ab63d649127.png"},{"id":60811113,"identity":"b3c95739-0901-40f3-b8ef-a66a151c09c3","added_by":"auto","created_at":"2024-07-22 10:53:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":669292,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTemporal consistency at the site of extensive attenuation.\u003c/strong\u003e(A) The distances of the lumen (yellow circles) and the EEM (red circles) from the center along an axis of 315 degrees. Through the consecutive frames at 0.2-mm intervals within the segment with the arc of IVUS signal loss \u0026gt; 90 degrees, the frame-to-frame variabilities for the lumen and EEM segmentation were assessed by the standard deviation of the distances (SD), and the standard deviation of the differences in the distances between two adjacent frames (SD-DIFF). (B) In a case example, both SD and SD-DIFF were much smaller in the model- (vs. expert-) derived EEM segmentation at the sites of severe attenuation.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4633591/v1/c8077adeee2ec5df3d0fcf67.png"},{"id":60811114,"identity":"0e83121e-7ab9-4bc0-beab-8b39f9f74b55","added_by":"auto","created_at":"2024-07-22 10:53:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":226817,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCorrelations between model- vs. expert-derived IVUS measurements and Bland-Altman plots.\u003c/strong\u003e(A \u0026amp; B) minimal lumen area, (C \u0026amp; D) lumen volume, (E \u0026amp; F) EEM volume, and (G \u0026amp; H) percent atheroma volume.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4633591/v1/a764ecdc8011b1107328838d.png"},{"id":63820769,"identity":"3af1da01-c2f4-43b3-afbe-6c9524dcfd72","added_by":"auto","created_at":"2024-09-02 16:07:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2165444,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4633591/v1/2ac25c8f-21b2-4131-ab69-34dbb6288e8a.pdf"},{"id":60812098,"identity":"0b27dca4-9894-4ca5-808d-24cb93bef928","added_by":"auto","created_at":"2024-07-22 11:01:21","extension":"doc","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":4176384,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplemental Figure 1. Architecture of ResNet-50 \u0026amp; Transformer.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplemental Figure 2. Bland-Altman plots between inter-observers of IVUS measurements.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplemental Figure 3. Correlations between inter-observer variabilities of IVUS measurements.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplemental Figure 4. Kaplan-Meier curves for event-free survival.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"segmentationsupplIJCImaging.doc","url":"https://assets-eu.researchsquare.com/files/rs-4633591/v1/82c9528ed74ea2ec0493e0a7.doc"}],"financialInterests":"Competing interest reported. Kim H \u0026 Lee G is an employee of Mediwhale Inc., Seoul, Korea. Min D is an employee of and Ingradient Inc., Seoul, Korea. Other authors report no conflicts of interest regarding this manuscript.","formattedTitle":"Development and Validation of Deep Learning Model for Intravascular Ultrasound Image Segmentation","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIntravascular ultrasound (IVUS) is a generalized intracoronary imaging modality that provides morphological information to the operators about the severity of stenosis, atheroma burden and plaque characteristics. IVUS is also useful for device sizing and stent optimization by correcting for suboptimal findings such as stent underexpansion, edge problems and procedure-related complications, which help prevent early and late stent failure.\u003csup\u003e1\u0026ndash;3\u003c/sup\u003e Although the clinical impact of IVUS-guided percutaneous coronary intervention (PCI) has been well established,\u003csup\u003e4,5\u003c/sup\u003e IVUS still remains underused. Interpretation and measurement of IVUS images is time-consuming and demands adequately trained personnel, which partly explains the low penetration rate of IVUS in the real world.\u003c/p\u003e \u003cp\u003ePrecise delineation of the lumen and external elastic membrane (EEM) borders is essential for a morphological assessment. However, frame-by-frame manual annotation is an error-prone task with high inter- and intra-observer variabilities. Especially in the presence of blood speckling obscuring the luminal border, tissue attenuation by intra-plaque calcium or lipid, or various image artifacts, it is tricky to extract the contour without consideration of the morphologic characteristics of the adjacent segments. Even with the selection of every 60th (1-mm interval) or 30th (0.5-mm interval) frame, volumetric analysis in an entire IVUS pullback containing thousands of images usually takes several hours. When derived from one selected image, the traditional IVUS criteria such as minimal lumen or stent area and plaque burden may be insufficient to represent the status of a whole vascular segment. Therefore, a program for automatic segmentation of the entire sequence of pullback images is necessary to facilitate the on-site utilization of IVUS and to stratify future cardiovascular risks.\u003c/p\u003e \u003cp\u003eAlthough a variety of methodologies for semantic segmentation have been applied, programs for automated IVUS analysis have not been widely employed in real practice, which is possibly due to their suboptimal performance, relatively small number of training samples, and incomplete clinical validation.\u003csup\u003e6\u0026ndash;11\u003c/sup\u003e It is uncertain if the models based on the 40-MHz IVUS images of native coronary arteries can be applied to 60-MHz IVUS images or stented lesions. Moreover, a lack of external validation limits the generalization of the developed programs. Thus, the aims of the current study were to 1) develop a 40-MHz IVUS-based deep learning model for delineating the lumen and EEM borders of native coronary arteries, 2) validate the performance and clinical impact of the model, and 3) evaluate whether the model can be applied to 60-MHz IVUS images or stented segments.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cstrong\u003eStudy population.\u0026nbsp;\u003c/strong\u003eBetween January 2014 and July 2016, pre-procedural IVUS was conducted in 1657 stable and unstable angina patients with an angiographic diameter stenosis \u0026gt;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.\u0026nbsp;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.\u003c/p\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcquisition of IVUS.\u003c/strong\u003e 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.\u0026nbsp;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 ≥\u0026nbsp;0.5 mm).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel development.\u003c/strong\u003e 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.\u003c/p\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003eEach 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.\u003c/p\u003e\n\u003cp\u003eTo assess the model performance, the extent of overlap between the\u0026nbsp;model-derived\u0026nbsp;vs. the expert-measured lumen and\u0026nbsp;EEM areas was assessed by three evaluation metrics (described in the Supplementary Appendix) including the Dice similarity\u0026nbsp;coefficient (DSC), Jaccard index (JI) and Surface Dice similarity coefficient (SDSC).\u0026nbsp;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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel validation.\u003c/strong\u003e 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.\u003csup\u003e13\u003c/sup\u003e Using\u0026nbsp;111 pre-procedural IVUS pullbacks, the intra- and inter-observer variances in the core laboratory analysis were assessed by Expert 1 and 2.\u003c/p\u003e\n\u003cp\u003eIn 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 \u0026gt; 90˚ without an ultrasound signal behind a lipid-rich plaque or calcification.\u003csup\u003e14-16\u003c/sup\u003e 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.\u003csup\u003e17\u003c/sup\u003e The extent of overlap between the\u0026nbsp;model-derived\u0026nbsp;vs. the expert-measured lumen and\u0026nbsp;EEM areas was assessed at the frame-level.\u003c/p\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003eIn addition, the model was applied to 60-MHz pre-procedural IVUS images\u0026nbsp;(OptiCross HD, Boston Scientific Corporation, Marlborough, Massachusetts, USA)\u0026nbsp;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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Validation.\u003c/strong\u003e 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 \u0026gt; 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\u0026nbsp;≥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).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExternal validation.\u003c/strong\u003e In 65 patients undergoing PCI\u0026nbsp;between April 2022 and July 2023, a total of\u0026nbsp;65 pre-stenting IVUS pullbacks\u0026nbsp;obtained by a\u0026nbsp;45-MHz IVUS catheter (Refinity, Philips Volcano, San Diego, CA, USA) were collected from Chung-Ang University Hospital,\u0026nbsp;Seoul, Republic of Korea. The\u0026nbsp;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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analysis.\u003c/strong\u003e 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 \u0026lt;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.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eClinical and lesion characteristics.\u0026nbsp;\u003c/strong\u003eIn the study cohort for model development, the mean age was 64.3\u0026nbsp;± 9.5 years, and 75% were men. The target vessels were the left anterior descending artery in\u0026nbsp;73%, the\u0026nbsp;left circumflex artery\u0026nbsp;in\u0026nbsp;5%, the right coronary artery in 18%, the ramus intermedius in 2%, and the left main coronary artery in 2%.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFrame-level performance.\u003c/strong\u003e Based on the evaluation metrics, the\u0026nbsp;frame-level performance for lumen and EEM segmentation are summarized in Table 1.\u0026nbsp;The model performance was also shown with the inclusion of 3,442 (17.8%) frames that showed an arc of IVUS signal loss \u0026gt; 90˚ behind attenuated or calcified plaque. Figure 2 compares the frame-to-frame variabilities of the model- vs. the expert-derived segmentation.\u003c/p\u003e\n\u003cp\u003eTo exclude the potential clustering effect of multiple frames in a vessel, the mean performance indices calculated in each of the 127 IVUS pullbacks in the test set were averaged. The averages of the mean DSC were 0.967 ± 0.007 and 0.982 ± 0.006 for the lumen and EEM segmentation, respectively. The averages of the mean JI were 0.938 ± 0.013 and 0.966 ± 0.011, respectively.\u0026nbsp;In addition, the averages of the mean SDSC were 0.869 ± 0.055 and 0.910 ± 0.011, respectively.\u003c/p\u003e\n\u003cp\u003eThe inference time was 0.029 second per frame, and the model-derived segmentation for 50-mm length (including 3000 frames) required 88 seconds. In comparison, manual segmentation with 0.2-mm interval took 187.5 minutes on average.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVessel-level performance.\u0026nbsp;\u003c/strong\u003eThe vessel-level performance of the model was evaluated in the cohort of the previous trial. The target vessels were the left anterior descending artery in 39%, the left circumflex artery in 19%, the right coronary artery in 40%, and the ramus intermedius in 2%. The lesion length was 25.9±8.7 mm, and plaque burden at the minimal lumen area (MLA) site was 67.6±10.9. There were no significant differences in the IVUS parameters measured by the model vs. the expert (Table 2). Overall the agreement between the model-derived vs. the expert-measured parameters was excellent (Figure 3). However, the mean MLA derived by the model (vs. measured by the expert) was significantly smaller (4.1±1.8 mm\u003csup\u003e2\u003c/sup\u003e vs. 4.3±2.0 mm\u003csup\u003e2\u003c/sup\u003e, p\u0026lt;0.001). The model-derived MLA was smaller than that measured by the expert in 78.8% of the cases. The intra- and inter-observer variations in the core laboratory analysis are shown in Supplemental Table 1, Supplemental Figure 2 and Supplemental Figure 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Validation.\u0026nbsp;\u003c/strong\u003eIn the retrospective cohort including 652 patients, baseline characteristics of patients and lesions are summarized in Supplemental Table 2.\u0026nbsp;The median follow-up was 47.8 months (IQR 40.1–67.8 months). At 3 years, cardiac death and acute MI occurred in 21 (3.2%) and 5 (0.8%) patients, respectively. Noncuprit-related TVR was performed in 28 (4.3%) patients during 3-year follow-up.\u003c/p\u003e\n\u003cp\u003eTable 3 compares the model-derived IVUS measurements between patients with vs. without 3-year cardiac death and TVR. There were no significant differences in the IVUS measurements or FFR between patients with vs. without acute MI. Among the IVUS variables, PAV\u0026gt;52.5% (area under curve 0.70, sensitivity 71% and specificity 67%) and plaque burden at the MLA site (area under curve 0.72, sensitivity 72% and specificity 66%) best predicted 3-year cardiac death and TVR, respectively (Supplemental Table 3). On the Kaplan-Meier curves (Supplemental Figures 4), the cardiac death-free survival rate at 3 years was significantly lower with PAV \u0026gt; 52.5% vs. ≤ 52.5% (93.4% vs. 98.6%, log-rank p \u0026lt; 0.001). In addition, the TVR-free survival rate was much lower with plaque burden at the MLA site \u0026gt;76.5% vs. ≤ 76.5% (91.5% vs. 98.1%, log-rank p \u0026lt;0.001).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStented segment.\u0026nbsp;\u003c/strong\u003eTable 4 shows the frame-level performances in the stented segments before and after fine-tuning of the model. For lumen segmentation, the DSC was increased from 0.918 ± 0.025 pre-tuning to\u0026nbsp;0.975 ± 0.010 post-tuning in the test set. To contour the EEM border, the DSC was improved from\u0026nbsp;0.928 ± 0.059 to\u0026nbsp;0.963 ± 0.026 by fine-tuning the model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e60-MHz IVUS images.\u0026nbsp;\u003c/strong\u003eWhen the model was tested on the 60-MHz IVUS images, the DSCs were 0.978 ± 0.024 and 0.985 ± 0.017 for the lumen and EEM, respectively. The JIs were 0.958 ± 0.040 and 0.972 ± 0.030, respectively.\u0026nbsp;The SDSCs were 0.970 ± 0.086 and 0.967 ± 0.098, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExternal validation.\u0026nbsp;\u003c/strong\u003eThe model was also tested in the external data. For lumen and EEM segmentation, the DSCs were 0.960 ± 0.049 and 0.979 ± 0.021, respectively. The JIs were 0.927 ± 0.078 and 0.960 ± 0.038, respectively. The SDSCs were 0.875 ± 0.181 and 0.892 ± 0.190, respectively.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eAlthough a variety of methodologies for automatic IVUS segmentation have been applied, there is a lack of models that are widely used in real practice. Their suboptimal performance can be explained by several reasons. With the requirement of a large data set for supervised deep learning, an insufficient number of manually-labeled training samples is one of the reasons for their poor accuracy. Moreover, there are fundamental pitfalls of IVUS images, including their limited spatial resolution, to-and-fro motion of the imaging catheter and various artifacts that make it challenging to delineate the vascular geometry. Especially at sites of tissue attenuation behind calcium or lipid, it has been considered that there is a need to utilize multiple adjacent frames for contouring the presumptive vessel border. However, it remains uncertain how to explicitly exploit a complex series of cross-sectional images.\u003c/p\u003e\n\u003cp\u003eAs a data-driven approach, convolutional neural networks that have been designed to automatically and adaptively ascertain the spatial hierarchies of features and have been adopted in many computer vision applications.\u003csup\u003e6-11\u003c/sup\u003e For IVUS segmentation, Yang et al. previously demonstrated that Dual Path U-Net had superior performance (JI 0.823 and 0.775 for lumen and EEM segmentation, respectively) relative to conventional computer vision-based approaches.\u003csup\u003e9\u003c/sup\u003e Nishi et al also developed DeepLabv3 with a modification of the encoder component.\u003csup\u003e10\u003c/sup\u003e Due to their exclusion of images containing tissue attenuation\u0026nbsp;≥90° of arc, image artifacts and large side branches, their model could not be applied to general use. By tracing the 4 longitudinal cutting planes spaced by 45 degrees (as the ground truths), Ziemer et al.\u0026nbsp;developed a model for IVUS segmentation at the dataset level.\u003csup\u003e11\u003c/sup\u003e In the presence of a saw-tooth artifact throughout the longitudinal view, electrocardiography-gated selection of the end-diastolic frames is mandatory, which raises concerns about inadequate temporal resolution. In addition, electrocardiogram-synchronized images are currently available only for 20-MHz IVUS system.\u003c/p\u003e\n\u003cp\u003eIn this study, we proposed a deep learning model combining ResNet-50 with a Transformer architecture which is adept at capturing long-range dependencies. This combination facilitates the interaction of extracted features across both spatial and temporal dimensions. Additionally, the implementation of a polar coordinate transformation aids in maintaining the integrity of the segmentation masks. With a dataset comprising 191,407 frames that were manually labeled, our model significantly advanced both frame- and vessel-level performance. Even at sites of extensive calcification or tissue attenuation, our model showed excellent performance (DSC \u0026gt; 0.96)\u0026nbsp;for contouring both the lumen and the EEM. Frame-by-frame\u0026nbsp;manual\u0026nbsp;delineation of the invisible EEM with a consistent criterion is challenging, which might lead to high frame-to-frame variability. However, this\u0026nbsp;current\u0026nbsp;model achieved a much better temporal consistency of EEM segmentation by incorporating the overall geometrical information from the series of adjacent frames.\u003c/p\u003e\n\u003cp\u003eAlthough the precise measurement of MLA is important to assess lesion severity, determine device size, and predict future clinical outcomes,\u003csup\u003e17,18\u003c/sup\u003e there is large variability in the selection of the MLA frame by visually inspecting the pullback. In our study, the model- (vs. the expert-) derived MLA was smaller in most cases. The model-based whole-frame analysis enables us to select the right frame of the MLA.\u0026nbsp;By eliminating the effect of human subjectivity or uncertainty of interpretation, the automatic analysis may be useful for the accurate diagnosis.\u003c/p\u003e\n\u003cp\u003eOur study validated clinical impact of the model-derived measurements. Among the IVUS parameters,\u0026nbsp;PAV\u0026nbsp;\u0026gt;52.5% and plaque burden at the MLA site \u0026gt;76.5%\u0026nbsp;best predicted 3-year cardiac death and nonculprit-related\u0026nbsp;TVR, respectively. The occurrence of cardiac death may depend on the overall atheroma burden in the entire vascular segment, while TVR is more likely determined by the localized disease status.\u003c/p\u003e\n\u003cp\u003eBecause metallic strut shadows usually interfere with the full visualization of the EEM, the application of the model derived from native coronary arteries to stented vessels caused the performance degradation. Nonetheless, fine-tuning by using a data set of stent images improved the performance for contouring both stent (DSC 0.975 ± 0.010) and the EEM (DSC 0.963 ± 0.026). Moreover, the\u0026nbsp;model based on the 40-MHz IVUS images consistently showed high accuracies for contouring the 60-MHz images.\u003c/p\u003e\n\u003cp\u003eExternal validation of a deep learning model is an essential step to determine its reproducibility and generalizability to different devices and settings, and to patients with diverse lesion characteristics. Applied to the 45-MHz IVUS images in the external PCI cohort, good performance of our model was demonstrated.\u0026nbsp;By saving time and expenses in core laboratories, the program potentially facilitates the on-site utilization of IVUS and real-time decision-making during PCI.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy limitations.\u003c/strong\u003e First, the current model needs to be validated in all of the commercially available systems, and tested on images obtained by different pullback speed or manual pullback. Second, even with its good performance for stent segmentation, this model cannot be used for assessing in-stent-restenosis. For a complete post-stenting evaluation, further tuning of the model should be performed for the meticulous delineation of neointima, tissue prolapse, stent malapposition and intimal dissection. In addition, model training in lesion subsets with intimal disruption, dissection, intraluminal thrombus and nodular calcification is necessary. With a lack of expert consensus for the segmentation of the complex bifurcation geometry, it was challenging to perform data labeling with incoherent criteria, which might be responsible for the suboptimal performance within the POC of bifurcation. Moreover, further study to evaluate possible ethnic differences in the model’s performance is required. Even though there is a concern about the reproducibility of measurement potentially affecting the quality of data labeling, the intra- and inter-observer variances in our core laboratory analysis were not considerable.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis deep learning-based model precisely delineated vascular geometry and improved temporal consistency. With a good performance, it can be applied to the stented segment and 60-MHz IVUS images. The model-derived IVUS measurements based on the whole-frame analysis had prognostic implication for the occurrence of cardiac mortality and TVR at long term.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eInformed consent and patient details.\u0026nbsp;\u003c/strong\u003eThe authors declare that this report does not contain any personal information that could lead to the identification of the patients.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions.\u0026nbsp;\u003c/strong\u003eAll authors attest that they meet the current International Committee of Medical Journal Editors (ICMJE) criteria for Authorship.\u003c/p\u003e\n\u003cp\u003eHyeonmin Kim: model development, data analysis, writing and edit the manuscript.\u003c/p\u003e\n\u003cp\u003eJune-Goo Lee and Gyu-Jun Jeong: methodology, supervision and review\u003c/p\u003e\n\u003cp\u003eGeunyoung Lee, H Cho and H Min: development of model, data processing, software and methodology\u003c/p\u003e\n\u003cp\u003eDaegyu Min: methodology, review and edit the manuscript.\u003c/p\u003e\n\u003cp\u003eSW Lee, Jun Hwan Cho, and Sungsoo Cho: validation, review and edit the manuscript.\u003c/p\u003e\n\u003cp\u003eSJ Kang: Conceptualization and design of study, data curation, methodology, supervision, validation, Funding, development of model, and writing and edit the original draft.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Competing interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eKim H \u0026amp; Lee G is an employee of Mediwhale Inc., Seoul, Korea. Min D is an employee of and\u0026nbsp;Ingradient Inc., Seoul, Korea. Other authors report no conflicts of interest regarding this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding support and author disclosures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by grants from the Ministry of Science and ICT (NRF-2021R1A2C2006831) and the Asan Institute for Life Sciences, Asan Medical Center, Seoul, Republic of Korea (2021IP0071-1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank\u0026nbsp;Mediwhale Inc. and\u0026nbsp;Ingradient Inc.\u0026nbsp;for their technical support.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statemen.\u0026nbsp;\u003c/strong\u003eThe datasets generated and/or analyzed during the current study are not publicly available because permission of sharing patient data was not granted by the Institutional Review Board but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eFujii K, Carlier SG, Mintz GS, Yang YM, Moussa I, Weisz G, Dangas G, Mehran R, Lansky AJ, Kreps EM, Collins M, Stone GW, Moses JW, Leon MB (2005) Stent underexpansion and residual reference segment stenosis are related to stent thrombosis after sirolimus-eluting stent implantation: an intravascular ultrasound study. J Am Coll Cardiol 45(7):995-998.\u003c/li\u003e\n \u003cli\u003eOkabe T, Mintz GS, Buch AN, Roy P, Hong YJ, Smith KA, Torguson R, Gevorkian N, Xue Z, Satler LF, Kent KM, Pichard AD, Weissman NJ, Waksman R (2007) Intravascular ultrasound parameters associated with stent thrombosis after drug-eluting stent deployment. Am J Cardiol 100(4):615-620.\u003c/li\u003e\n \u003cli\u003eLiu X, Doi H, Maehara A, Mintz GS, Costa Jde R Jr, Sano K, Weisz G, Dangas GD, Lansky AJ, Kreps EM, Collins M, Fahy M, Stone GW, Moses JW, Leon MB, Mehran R. A (2009) A volumetric intravascular ultrasound comparison of early drug-eluting stent thrombosis versus restenosis. JACC Cardiovasc Interv 2(5): 428-434.\u003c/li\u003e\n \u003cli\u003eLee JM, Choi KH, Song YB, Lee JY, Lee SJ, Lee SY, Kim SM, Yun KH, Cho JY, Kim CJ, Ahn HS, Nam CW, Yoon HJ, Park YH, Lee WS, Jeong JO, Song PS, Doh JH, Jo SH, Yoon CH, Kang MG, Koh JS, Lee KY, Lim YH, Cho YH, Cho JM, Jang WJ, Chun KJ, Hong D, Park TK, Yang JH, Choi SH, Gwon HC, Hahn JY; RENOVATE-COMPLEX-PCI Investigators (2023) Intravascular Imaging-Guided or Angiography-Guided Complex PCI. N Engl J Med 388(18):1668-1679.\u003c/li\u003e\n \u003cli\u003eWitzenbichler B, Maehara A, Weisz G, Neumann FJ, Rinaldi MJ, Metzger DC, Henry TD, Cox DA, Duffy PL, Brodie BR, Stuckey TD, Mazzaferri EL Jr, Xu K, Parise H, Mehran R, Mintz GS, Stone GW (2014) Relationship between intravascular ultrasound guidance and clinical outcomes after drug-eluting stents: the assessment of dual antiplatelet therapy with drug-eluting stents (ADAPT-DES) study. Circulation 129(4):463-470.\u003c/li\u003e\n \u003cli\u003eBajaj R, Huang X, Kilic Y, Ramasamy A, Jain A, Ozkor M, Tufaro V, Safi H, Erdogan E, Serruys PW, Moon J, Pugliese F, Mathur A, Torii R, Baumbach A, Dijkstra J, Zhang Q, Bourantas CV (2021) Advanced deep learning methodology for accurate, real-time segmentation of high-resolution intravascular ultrasound images. Int J Cardiol 339:185-191.\u003c/li\u003e\n \u003cli\u003eZhu F, Gao Z, Zhao C, Zhu H, Nan J, Tian Y, Dong Y, Jiang J, Feng X, Dai N, Zhou W (2022) A Deep Learning-based Method to Extract Lumen and Media-Adventitia in Intravascular Ultrasound Images. Ultrason Imaging 44(5-6):191-203.\u003c/li\u003e\n \u003cli\u003eMendizabal-Ruiz EG, Rivera M, Kakadiaris IA (2013) Segmentation of the luminal border in intravascular ultrasound B-mode images using a probabilistic approach. Med Image Anal 17(6):649-670.\u003c/li\u003e\n \u003cli\u003eYang J, Faraji M, Basu A (2019) Robust segmentation of arterial walls in intravascular ultrasound images using Dual Path U-Net. Ultrasonics 96:24-33.\u003c/li\u003e\n \u003cli\u003eNishi T, Yamashita R, Imura S, Tateishi K, Kitahara H, Kobayashi Y, Yock PG, Fitzgerald PJ, Honda Y (2021) Deep learning-based intravascular ultrasound segmentation for the assessment of coronary artery disease. Int J Cardiol 333:55-59.\u003c/li\u003e\n \u003cli\u003eZiemer PGP, Bulant CA, Orlando JI, Maso Talou GD, \u0026Aacute;lvarez LAM, Guedes Bezerra C, Lemos PA, Garc\u0026iacute;a-Garc\u0026iacute;a HM, Blanco PJ (2020) Automated lumen segmentation using multi-frame convolutional neural networks in intravascular ultrasound datasets. Eur Heart J Digit Health 1(1):75-82.\u003c/li\u003e\n \u003cli\u003ePark SJ, Kang SJ, Ahn JM, Chang M, Yun SC, Roh JH, Lee PH, Park HW, Yoon SH, Park DW, Lee SW, Kim YH, Lee CW, Mintz GS, Han KH, Park SW (2016) Effect of Statin Treatment on Modifying Plaque Composition: A Double-Blind, Randomized Study. J Am Coll Cardiol 67(15):1772-1783.\u003c/li\u003e\n \u003cli\u003eMintz GS, Nissen SE, Anderson WD, Bailey SR, Erbel R, Fitzgerald PJ, Pinto FJ, Rosenfield K, Siegel RJ, Tuzcu EM, Yock PG (2001) American College of Cardiology Clinical Expert Consensus Document on Standards for Acquisition, Measurement and Reporting of Intravascular Ultrasound Studies (IVUS): a report of the American College of Cardiology Task Force on Clinical Expert Consensus Documents. J Am Coll Cardiol 37(5):1478-1492.\u003c/li\u003e\n \u003cli\u003eKang SJ, Mintz GS, Pu J, Sum ST, Madden SP, Burke AP, Xu K, Goldstein JA, Stone GW, Muller JE, Virmani R, Maehara A (2015) Combined IVUS and NIRS detection of fibroatheromas: histopathological validation in human coronary arteries. JACC Cardiovasc Imaging 8(2):184-194.\u003c/li\u003e\n \u003cli\u003ePu J, Mintz GS, Biro S, Lee JB, Sum ST, Madden SP, Burke AP, Zhang P, He B, Goldstein JA, Stone GW, Muller JE, Virmani R, Maehara A (2014) Insights into echo-attenuated plaques, echolucent plaques, and plaques with spotty calcification: novel findings from comparisons among intravascular ultrasound, near-infrared spectroscopy, and pathological histology in 2,294 human coronary artery segments. J Am Coll Cardiol 63(21):2220-2233.\u003c/li\u003e\n \u003cli\u003eKang SJ, Ahn JM, Han S, Park DW, Lee SW, Kim YH, Lee CW, Park SW, Mintz GS, Park SJ (2016) Multimodality imaging of attenuated plaque using grayscale and virtual histology intravascular ultrasound and optical coherent tomography. Catheter Cardiovasc Interv 88(1):E1-11.\u003c/li\u003e\n \u003cli\u003eKang SJ, Mintz GS, Oh JH, Park DW, Lee SW, Kim YH, Lee CW, Park SW, Park SJ (2013) Intravascular ultrasound assessment of distal left main bifurcation disease: the importance of the polygon of confluence of the left main, left anterior descending, and left circumflex arteries. Catheter Cardiovasc Interv 82(5):737-745.\u003c/li\u003e\n \u003cli\u003eKang SJ, Lee JY, Ahn JM, Mintz GS, Kim WJ, Park DW, Yun SC, Lee SW, Kim YH, Lee CW, Park SW, Park SJ (2011) Validation of intravascular ultrasound-derived parameters with fractional flow reserve for assessment of coronary stenosis severity. Circ Cardiovasc Interv 4(1):65-71.\u003c/li\u003e\n \u003cli\u003eStone GW, Maehara A, Lansky AJ, de Bruyne B, Cristea E, Mintz GS, Mehran R, McPherson J, Farhat N, Marso SP, Parise H, Templin B, White R, Zhang Z, Serruys PW; PROSPECT Investigators (2011) A prospective natural-history study of coronary atherosclerosis. N Engl J Med 364(3);226-235.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;Frame-level performance for lumen and EEM segmentation\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"917\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd colspan=\"3\"\u003e\n \u003cp\u003eFor lumen segmentation*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\"\u003e\n \u003cp\u003e\u0026nbsp;For\u0026nbsp;EEM segmentation*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eJI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eJI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eOverall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eValidation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.966 \u0026plusmn; 0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.936 \u0026plusmn; 0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.871 \u0026plusmn; 0.142\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.982 \u0026plusmn; 0.017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.965 \u0026plusmn; 0.029\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.911 \u0026plusmn; 0.134\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTest set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.968 \u0026plusmn; 0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.938 \u0026plusmn; 0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.874 \u0026plusmn; 0.145\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.982 \u0026plusmn; 0.017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.966 \u0026plusmn; 0.031\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.910 \u0026plusmn; 0.141\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eSegments with extensive IVUS attenuation\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.968 \u0026plusmn; 0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.938 \u0026plusmn; 0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.889 \u0026plusmn; 0.136\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.973 \u0026plusmn; 0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.950 \u0026plusmn; 0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.803 \u0026plusmn; 0.191\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWithin the POC\u003csup\u003e\u0026Dagger;\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.941 \u0026plusmn; 0.039\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.892 \u0026plusmn; 0.067\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.747 \u0026plusmn; 0.147\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.968 \u0026plusmn; 0.026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.940 \u0026plusmn; 0.046\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.808 \u0026plusmn; 0.125\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eEEM = external elastic membrane,\u0026nbsp;DSC = Dice similarity coefficient, JI = Jaccard index, SDSC = Surface Dice similarity coefficient\u003c/p\u003e\n\u003cp\u003e* Including all frames at 0.2-mm (12 frames) intervals\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003csup\u003e\u0026nbsp;\u003c/sup\u003eIncluding 123 segments with 3442 frames (in the test set) that\u0026nbsp;showed an arc of IVUS signal loss \u0026gt; 90˚\u0026nbsp;behind the attenuated or calcified plaque\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e\u0026Dagger;\u0026nbsp;\u003c/sup\u003e206 bifurcation segments defined as the polygon of confluence between the carina and the distal end of the proximal main branch\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e \u003cstrong\u003eComparison of model- vs. expert-derived measurements.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"566\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\"\u003e\n \u003cp\u003eExpert\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eCross-sectional\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eLesion length\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e25.9\u0026plusmn;8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e25.8\u0026plusmn;8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eMinimal lumen area, mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e4.3\u0026plusmn;2.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e4.1\u0026plusmn;1.8*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eEEM at the MLA site, mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e13.6\u0026plusmn;4.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e13.3\u0026plusmn;4.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eP+M at the MLA site, mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e9.3\u0026plusmn;3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e9.2\u0026plusmn;3.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003ePlaque burden at the MLA site, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e67.6\u0026plusmn;10.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e68.6\u0026plusmn;10.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eMean EEM diameter, mm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e4.1\u0026plusmn;0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e4.1\u0026plusmn;0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eVolumetric\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eLumen volume, mm\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e194.0\u0026plusmn;92.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e195.3\u0026plusmn;92.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003eEEM volume, mm\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e401.6\u0026plusmn;186.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e402.8\u0026plusmn;183.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003ePlaque volume, mm\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e207.6\u0026plusmn;107.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e207.5\u0026plusmn;103.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"51.9434628975265%\"\u003e\n \u003cp\u003ePercent atheroma volume (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e51.3\u0026plusmn;8.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.02826855123675%\" valign=\"top\"\u003e\n \u003cp\u003e51.2\u0026plusmn;7.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*p values \u0026lt;0.05 (vs. expert)\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Model-derived IVUS parameters and FFR between patients with and without 3-year clinical events.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"926\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.7170626349892%\" colspan=\"2\"\u003e\n \u003cp\u003eCardiac death\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"35.745140388768895%\" colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eTarget vessel revascularization\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\"\u003e\n \u003cp\u003e(+)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\"\u003e\n \u003cp\u003e(-)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e(+)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e(-)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003e\u003cstrong\u003eIVUS measurements\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eMLA, mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e2.7 (2.0 \u0026ndash; 4.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e2.9 (2.2 \u0026ndash; 3.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e2.3 (2.0 \u0026ndash; 3.0 )\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e2.9 (2.2 \u0026ndash; 3.7)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eEEM area at the MLA site, mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e9.6 (7.1 \u0026ndash; 14.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e10.9 (8.0 \u0026ndash; 14.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e11.6 (9.9 \u0026ndash; 15.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e10.9 (7.8 \u0026ndash; 14.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eP+M area at the MLA site ,\u0026nbsp;mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e7.1 (5.1 \u0026ndash; 10.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e8.0 (5.2 \u0026ndash; 10.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e9.4 (7.5 \u0026ndash; 12.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e7.8 (5.1 \u0026ndash; 10.7)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003ePlaque burden at the MLA site, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e75.9 (69.1 \u0026ndash; 77.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e73.4 (65.8 \u0026ndash; 78.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e80.1 (72.1 \u0026ndash; 83.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e73.3 (65.8 \u0026ndash; 78.4)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eLumen volume, mm\u003csup\u003e3\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e269.1 (220.9 \u0026ndash; 489.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e310.6 (208.8 \u0026ndash; 416.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e247.0 (153.1 \u0026ndash; 350.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e312.2 (213.4 \u0026ndash; 426.4)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eEEM volume, mm\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e614.4 (474.9 \u0026ndash; 1106.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e609.1 (421.4 \u0026ndash; 819.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e518.4 (365.3 \u0026ndash; 725.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e612.7 (433.9 \u0026ndash; 826.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eP+M volume, mm\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e382.4 (249.6 \u0026ndash; 587.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e290.8 (206.9 \u0026ndash; 401.7)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e232.5 (199.1 \u0026ndash; 380.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e296.2 (208.7 \u0026ndash; 407.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003ePercent atheroma volume, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e55.7 (49.7 \u0026ndash; 58.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e48.8 (43.5 \u0026ndash; 54.2)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e54.8 (45.0 \u0026ndash; 59.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e48.8 (43.5 \u0026ndash; 54.2)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003eLength of ROI, mm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e53.2 (37.6 \u0026ndash; 65.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e45.7 (34.3 \u0026ndash; 56.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e41.4 (27.1 \u0026ndash; 51.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e45.9 (34.6 \u0026ndash; 56.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"27.5377969762419%\"\u003e\n \u003cp\u003e\u003cstrong\u003eFFR at maximal hyperemia\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e0.86 (0.83 \u0026ndash; 0.89)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e0.86 (0.81 \u0026ndash; 0.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.386609071274297%\" valign=\"top\"\u003e\n \u003cp\u003e0.81 (0.74 \u0026ndash; 0.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.3585313174946%\" valign=\"top\"\u003e\n \u003cp\u003e0.86 (0.82 \u0026ndash; 0.90)\u003csup\u003e#\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eMLA= minimal lumen area, EEM=external elastic membrane, P+M= plaque + media, ROI= region of interest, * p value \u0026lt; 0.05 vs. 3-year cardiac death (+) group (by Mann-Whitney), \u003csup\u003e#\u003c/sup\u003e p value \u0026lt; 0.05 vs. 3-year target vessel revascularization (+) group (by Mann-Whitney)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4.\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;Frame-level performance in stented segments\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"928\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd colspan=\"3\"\u003e\n \u003cp\u003eFor lumen segmentation*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\"\u003e\n \u003cp\u003e\u0026nbsp;For\u0026nbsp;EEM segmentation*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eJI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eJI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSDSC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePre-tuning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eValidation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.931\u0026plusmn;0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.872\u0026plusmn;0.039\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.371\u0026plusmn;0.208\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.949\u0026plusmn;0.051\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.908\u0026plusmn;0.084\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.621\u0026plusmn;0.292\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTest set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.918\u0026plusmn;0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.850\u0026plusmn;0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.281\u0026plusmn;0.179\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.928\u0026plusmn;0.059\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.870\u0026plusmn;0.096\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.458\u0026plusmn;0.307\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePost-tuning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eValidation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.976\u0026plusmn;0.010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.953\u0026plusmn;0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.875\u0026plusmn;0.116\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.971\u0026plusmn;0.026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.945\u0026plusmn;0.047\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.758\u0026plusmn;0.239\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTest set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.975\u0026plusmn;0.010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.952\u0026plusmn;0.018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.881\u0026plusmn;0.110\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.963\u0026plusmn;0.026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.930\u0026plusmn;0.046\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.680\u0026plusmn;0.239\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*Including all frames at 0.2-mm (12 frames) intervals\u003c/p\u003e\n\u003cp\u003eEEM = external elastic membrane, DSC = Dice similarity coefficient, JI = Jaccard index, SDSC = Surface Dice similarity coefficient\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-cardiovascular-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"caim","sideBox":"Learn more about [The International Journal of Cardiovascular Imaging](https://www.springer.com/journal/10554)","snPcode":"10554","submissionUrl":"https://submission.nature.com/new-submission/10554/3","title":"The International Journal of Cardiovascular Imaging","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"intravascular ultrasound, segmentation, deep learning, coronary artery disease ","lastPublishedDoi":"10.21203/rs.3.rs-4633591/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4633591/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eAims. \u003c/strong\u003eThis study was conducted to develop and validate a deep learning model for delineating intravascular ultrasound (IVUS) images of coronary arteries.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMaterials and Methods. \u003c/strong\u003eUsing atotal 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults.\u003c/strong\u003e 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 \u0026gt; 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 good in the independent retrospective cohort (all, intra-class coefficients \u0026gt; 0.94). The model-derived PAV\u0026gt;52.5% (area under curve 0.70, sensitivity 71% and specificity 67%) and plaque burden at the MLA 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 \u0026gt; 0.96 for contouring lumen and EEM were achieved by fine-tuning. Applied to the 60-MHz IVUS images, the DSCs were \u0026gt; 0.97. In the external cohort with 45-MHz IVUS, the DSCs were \u0026gt; 0.96.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion.\u003c/strong\u003e The deep learning model accurately delineated vascular geometry, which may be cost-saving and support clinical decision-making.\u003c/p\u003e","manuscriptTitle":"Development and Validation of Deep Learning Model for Intravascular Ultrasound Image Segmentation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-07-22 10:53:16","doi":"10.21203/rs.3.rs-4633591/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-07-12T15:41:31+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-12T11:57:33+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-11T15:38:59+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-11T14:55:44+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-10T16:51:51+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-06T02:28:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"325974919793598108912583227179456125426","date":"2024-07-01T20:21:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"205129598542182371169749241148234983427","date":"2024-06-28T18:38:43+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"302775699959370764199669916009527214818","date":"2024-06-28T14:41:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"207426154722468402827088586983619696458","date":"2024-06-28T13:19:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"168204387807834925595574480156210963210","date":"2024-06-28T12:24:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"31388213310105687356789156095739926419","date":"2024-06-28T12:22:40+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-06-28T12:18:12+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-06-25T10:28:07+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-06-25T10:27:37+00:00","index":"","fulltext":""},{"type":"submitted","content":"The International Journal of Cardiovascular Imaging","date":"2024-06-25T05:05:40+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-cardiovascular-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"caim","sideBox":"Learn more about [The International Journal of Cardiovascular Imaging](https://www.springer.com/journal/10554)","snPcode":"10554","submissionUrl":"https://submission.nature.com/new-submission/10554/3","title":"The International Journal of Cardiovascular Imaging","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"735037cf-b5d2-40d1-b05b-a35c5adf9303","owner":[],"postedDate":"July 22nd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2024-09-02T15:58:55+00:00","versionOfRecord":{"articleIdentity":"rs-4633591","link":"https://doi.org/10.1007/s10554-024-03221-9","journal":{"identity":"the-international-journal-of-cardiovascular-imaging","isVorOnly":false,"title":"The International Journal of Cardiovascular Imaging"},"publishedOn":"2024-08-27 15:56:56","publishedOnDateReadable":"August 27th, 2024"},"versionCreatedAt":"2024-07-22 10:53:16","video":"","vorDoi":"10.1007/s10554-024-03221-9","vorDoiUrl":"https://doi.org/10.1007/s10554-024-03221-9","workflowStages":[]},"version":"v1","identity":"rs-4633591","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4633591","identity":"rs-4633591","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}