Elsevier

International Journal of Cardiology

Volume 339, 15 September 2021, Pages 185-191
International Journal of Cardiology

Advanced deep learning methodology for accurate, real-time segmentation of high-resolution intravascular ultrasound images

https://doi.org/10.1016/j.ijcard.2021.06.030Get rights and content

Highlights

  • Novel machine learning methodology for IVUS segmentation trained and tested in the largest dataset to date.

  • The methodology introduced reliably matches human experts at segmenting IVUS images but is an order of magnitude faster.

  • In user-friendly software the methodology is expected to enable real-time IVUS analysis and optimal treatment planning.

Abstract

Aims

The aim of this study is to develop and validate a deep learning (DL) methodology capable of automated and accurate segmentation of intravascular ultrasound (IVUS) image sequences in real-time.

Methods and results

IVUS segmentation was performed by two experts who manually annotated the external elastic membrane (EEM) and lumen borders in the end-diastolic frames of 197 IVUS sequences portraying the native coronary arteries of 65 patients. The IVUS sequences of 177 randomly-selected vessels were used to train and optimise a novel DL model for the segmentation of IVUS images. Validation of the developed methodology was performed in 20 vessels using the estimations of two expert analysts as the reference standard.

The mean difference for the EEM, lumen and plaque area between the DL-methodology and the analysts was ≤0.23mm2 (standard deviation ≤0.85mm2), while the Hausdorff and mean distance differences for the EEM and lumen borders was ≤0.19 mm (standard deviation≤0.17 mm). The agreement between DL and experts was similar to experts' agreement (Williams Index ranges: 0.754–1.061) with similar results in frames portraying calcific plaques or side branches.

Conclusions

The developed DL-methodology appears accurate and capable of segmenting high-resolution real-world IVUS datasets. These features are expected to facilitate its broad adoption and enhance the applications of IVUS in clinical practice and research.

Introduction

Intravascular ultrasound (IVUS) is the preferred modality for assessing lumen and vessel wall dimensions, quantifying plaque burden and guiding revascularisation in complex lesions and high-risk patients. [1] Cumulative data have underscored its value in optimising procedural results and improving outcomes in patients undergoing percutaneous coronary intervention (PCI). [[2], [3], [4]] Its broader use, however, has been limited by the time and cost needed for acquiring and analysing IVUS imaging. Whilst several automated solutions have been proposed to aid IVUS segmentation, none has proven capable of processing IVUS images and accurately identifying the external elastic membrane (EEM) and lumen borders in real-time while the patient is in the cardiac catheterisation laboratory. [5]

Deep learning (DL) methodologies have begun to find medical application in facilitating the rapid and accurate processing of large imaging datasets. They rely on learning from large datasets of human annotations and use this information to train algorithmic models capable of processing images and replicating human performance within a few seconds. [6,7] These features render DL as the ideal approach for the analysis of IVUS images that have increased noise and artifacts which can confound conventional image-based segmentation approaches. Despite the undoubted theoretical advantages of DL methodologies in this setting, few studies have tested their potential value in IVUS analysis, mainly in small datasets providing promising results (Supplementary Table 1). [[8], [9], [10], [11], [12], [13], [14], [15], [16], [17]] The aim of this study is to develop and validate a DL-methodology capable of automatically and accurately segmenting IVUS data in real-time.

Section snippets

Study population

Seventy patients who provided written, informed consent were recruited to the “Evaluation of the efficacy of computed tomographic coronary angiography in assessing coronary artery morphology and physiology” study (NCT03556644) and included in this analysis. The rationale and the study design have been presented in detail previously. [18] All patients had stable angina due to atherosclerosis detected on cardiac catheterisation and were referred to Barts Heart Centre for further invasive

Results

NIRS-IVUS was performed in 65 (197 vessels; 3.03 vessels/patient) out of 70 recruited patients; one patient was excluded due to acute kidney injury post CTCA, one patient was excluded due to findings of a thymoma on CTCA and three cases due to NIRS-IVUS catheter dysfunction. The baseline demographics of the studied patients and the output of the IVUS analysis are shown in Supplementary Table 2 and 3 respectively.

The training set consisted of 824,750 frames (177 vessels), of which 23,774

Discussion

In the present study we developed a novel DL-methodology and examined its efficacy in detecting the EEM and lumen borders in high-resolution IVUS images. The methodology was trained in a large dataset of 177 vessels and tested in 20 vessels. Both sets included frames portraying common artifacts seen in IVUS (guide-wire artifacts, non-uniform rotational distortion, motion artifacts, reverberations, side lobe artifacts and blood speckle artifacts) as well as frames with calcific tissue and side

Conclusions

In the present study we present for the first time a novel DL-methodology that is capable of accurate, automated segmentation of high-resolution IVUS images in real-time. An extensive morphological and quantitative validation in a large dataset has demonstrated that the proposed method is able to overcome common artifacts seen in IVUS and segment challenging frames portraying the origin of side branches and calcific plaques. These features are expected to facilitate its broad adoption and

Sources of funding

This study is jointly funded by British Heart Foundation (PG/17/18/32883), University College London Biomedical Research Centre (BRC492B) and Rosetrees Trust (A1773). RB, AR, VT, AM, AB and CVB are funded by Barts NIHR Biomedical Research Centre, London, UK.

Declaration of Competing Interest

All authors have no conflicts of interests to declare.

Acknowledgements

The authors wish to acknowledge the Cardiovascular Devices Hub at the Centre for Cardiovascular Medicine and Devices, Queen Mary University of London for supporting the present study.

References (36)

Cited by (16)

  • Machine learning for atherosclerotic tissue component classification in combined near-infrared spectroscopy intravascular ultrasound imaging: Validation against histology

    2022, Atherosclerosis
    Citation Excerpt :

    The only contemporary, commercially available tool for OCT plaque composition analysis includes a machine learning algorithm trained against the estimations of experts but not histology, approximating the external elastic membrane border in diseased segments, raising a concern about the accuracy of this approach in advanced atherosclerotic plaques [48]. Our approach overcomes these limitations as it was trained using histological estimations, has been validated in a discrete, independent test set, and has been incorporated in a user-friendly, commercially-available software – which also includes machine learning algorithms for fast and accurate segmentation of the IVUS images enabling analysis of large datasets and quantification of plaque components in a reproducible manner [49,50]. These advances are anticipated to promote the broad use of the combined ECHO-NIRS approach for plaque characterization in serial NIRS-IVUS studies evaluating the efficacy of novel pharmacotherapies on plaque evolution.

View all citing articles on Scopus
1

The 1st and 2nd author contributed equally to this work.

View full text