Echocardiography-based AI detection of regional wall motion abnormalities and quantification of cardiac function in myocardial infarction

Xixiang Lin, Feifei Yang, Yixin Chen, Xiaotian Chen, Wenjun Wang, Xu Chen, Qiushuang Wang, Liwei Zhang, Huayuan Guo, Bohan Liu, Liheng Yu, Haitao Pu, Peifang Zhang, Zhenhua Wu, Xin Li, Daniel Burkhoff, Kunlun HeFrontiers in Cardiovascular Medicine 2022 · DOI

TL;DR

This paper introduces a novel deep learning (DL) framework that automates the analysis of echocardiograms for detecting regional wall motion abnormalities (RWMAs) and quantifying cardiac function in patients with suspected myocardial infarction (MI). The system demonstrates high accuracy on both standard and portable bedside ultrasound devices, offering a potential solution for rapid, objective, and accessible cardiac assessment in emergency settings.

What Problem Does This Paper Address?

Myocardial infarction (MI) is a severe condition requiring rapid diagnosis and intervention to minimize mortality and disability. Echocardiography, particularly bedside ultrasound, is a crucial non-invasive tool for assessing cardiac function and detecting RWMAs in patients presenting with chest pain. However, accurate interpretation of echocardiograms demands highly trained and experienced physicians, who are often in short supply, especially in emergency departments. Furthermore, image quality from portable bedside equipment can be lower than from standard machines, potentially leading to diagnostic errors. This paper addresses the need for an automated, objective, and efficient system that can reliably analyze echocardiograms from both standard and bedside devices, overcoming the limitations of human expertise and varying image quality.

Methods

The researchers developed a multi-stage deep learning pipeline:

  • View Selection: An Xception Net model automatically identifies the three standard apical views (apical 4-chamber, apical 2-chamber, and apical long axis) crucial for analysis.
  • Quality Control: An automated algorithm assesses image quality, excluding suboptimal images that could hinder accurate diagnosis. This mimics clinical practice by not pre-screening images based on strict criteria.
  • Segmentation: A pretrained LSTM-Unet model segments the left ventricle into three regions corresponding to coronary artery territories (anterior, inferior, and anteroseptal walls). This segmentation is also used to outline other cardiac chambers for function quantification.
  • RWMA Detection: A Deep 3D Convolutional Neural Network (CNN) analyzes the segmented video frames to detect regional wall motion abnormalities.
  • Cardiac Function Quantification: The model automatically measures key cardiac parameters like left ventricular ejection fraction (LVEF), end-diastolic volume (LVEDV), and end-systolic volume (LVESV) using the segmentation output and the biplane method of disks summation.

The pipeline is designed to be fully automated, processing echocardiograms from raw video to diagnostic insights.

Datasets

  • Training and Validation: 2,274 transthoracic echocardiographic examinations from the Fourth Medical Center of Chinese PLA General Hospital were used, split into an 8:2 ratio for training and validation. These included both standard and bedside ultrasound data.
  • Internal Test Dataset: 1,868 examinations from the same hospital were used for internal testing.
  • External Test Dataset: 3,026 examinations from the Second Affiliated Hospital of Shandong University of Traditional Chinese Medicine were used for external validation.

The datasets included both patients with MI and control cases, matched for age and sex. The study also involved a separate evaluation where 100 cases were randomly selected for assessment by experts and beginners, both with and without AI assistance.

Key Insights and Ideas

  • Bedside Ultrasound Equivalence: A critical finding is that the DL model performs comparably well on images from portable bedside ultrasound devices as it does on images from standard equipment. This is significant because bedside ultrasound is increasingly used in emergency settings, but its lower image quality has been a concern.
  • Automated Quality Control Mimics Clinical Practice: The approach of not pre-screening images for quality but rather having an automated quality control step that excludes uninterpretable images is a practical advantage. This reflects how physicians often encounter varied image quality in real-world scenarios.
  • Beginner Assistance: The AI tool significantly improved the accuracy of beginners in detecting RWMAs, suggesting its potential as an educational aid to enhance skills in image acquisition and interpretation.
  • Comprehensive Pipeline: The development of a fully automated pipeline, from view selection and quality control to segmentation, RWMA detection, and cardiac function quantification, represents a significant step towards a complete AI-driven echocardiography analysis solution.

Key Results

  • View Selection Accuracy: The view selection model achieved an average accuracy of 96%.
  • Segmentation Performance: The segmentation model showed good agreement with manual segmentation, achieving an average Dice coefficient of 0.89.
  • RWMA Detection Accuracy:
    • Internal Test Dataset: Average AUROC of 0.91 for standard equipment and 0.88 for bedside equipment.
  • External Test Dataset: Average AUROC of 0.90 for standard equipment and 0.85 for bedside equipment.
  • Cardiac Function Quantification: Automatic cardiac function measurements showed good agreement with echocardiographic reports, with a mean bias of 4% for LVEF.
  • Image Exclusion Rates: The model excluded 2.7% of standard equipment studies and 14.7% of portable bedside studies as unqualified, a rate considered more practical than some previous studies.

Limitations and Future Work

  • "Black Box" Problem: The authors acknowledge the inherent "black box" nature of deep learning models, where understanding the exact features driving predictions can be challenging. They mitigated this by focusing the segmentation on the left ventricle.
  • Severity vs. Presence of RWMAs: The model detects the presence of RWMAs but not their severity, as wall motion scoring can be subject to inter- and intra-observer variability.
  • Age-Related Accuracy: The study found that older patients were more likely to have incorrect AI predictions, potentially due to generally lower wall motion in this demographic.
  • Future Work: The authors suggest that prospective randomized controlled trials (RCTs) are warranted to further validate the pipeline's performance in real-world clinical settings. They also highlight the potential of the model as an educational tool for improving physician skills.

Summary updated

Abstract

We present a fully automated echocardiography pipeline applicable to both standard and bedside ultrasound with various functions, including view selection, quality control, segmentation, detection of the region of wall motion abnormalities and quantification of cardiac function.
BibTeX
@inproceedings{lin2022echocardiography,
  author = {Xixiang Lin and Feifei Yang and Yixin Chen and Xiaotian Chen and Wenjun Wang and Xu Chen and Qiushuang Wang and Liwei Zhang and Huayuan Guo and Bohan Liu and Liheng Yu and Haitao Pu and Peifang Zhang and Zhenhua Wu and Xin Li and Daniel Burkhoff and Kunlun He},
  title = {{Echocardiography-based AI detection of regional wall motion abnormalities and quantification of cardiac function in myocardial infarction}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2022},
  doi = {10.3389/fcvm.2022.903660},
}
Figure