Source-linked AI summary
High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy with Cardiovascular Deep Learning
Grant Duffy, Paul P Cheng, Neal Yuan, Bryan He, Alan C. Kwan, Matthew J. Shun-Shin, Kevin M. Alexander, Joseph Ebinger, Matthew P. Lungren, Florian Rader, David H. Liang, Ingela Schnittger, Euan A. Ashley, James Y. Zou, Jignesh Patel, Ronald Witteles, Susan Cheng, David Ouyang
TL;DR
LVH is difficult to measure precisely and to distinguish by etiology because different diseases can produce similar morphology. EchoNet-LVH addresses this gap with automated echocardiographic quantification and video-based etiology prediction, achieving expert-level ventricular assessment and classification performance across internal and external datasets.
Problem
LVH assessment is limited by measurement error and variability, under-recognition, and difficulty differentiating its etiologies.
Method
EchoNet-LVH is a deep learning system that automatically quantifies ventricular hypertrophy from echocardiography and predicts LVH etiology.
Results
EchoNet-LVH achieved state-of-the-art ventricular thickness and diameter assessment within clinical test-retest variance and detected subtle phenotypes through high-throughput measurements.
Takeaways & Limitations
The system provides automated, high-throughput echocardiographic phenotyping and supports further precision diagnosis of cardiac hypertrophy.
Takeaways & Limitations
Prospective deployment and testing of AI systems in diverse clinical environments remain necessary.
Abstract
from arXiv · showhide
Left ventricular hypertrophy (LVH) results from chronic remodeling caused by a broad range of systemic and cardiovascular disease including hypertension, aortic stenosis, hypertrophic cardiomyopathy, and cardiac amyloidosis. Early detection and characterization of LVH can significantly impact patient care but is limited by under-recognition of hypertrophy, measurement error and variability, and difficulty differentiating etiologies of LVH. To overcome this challenge, we present EchoNet-LVH - a deep learning workflow that automatically quantifies ventricular hypertrophy with precision equal to human experts and predicts etiology of LVH. Trained on 28,201 echocardiogram videos, our model accurately measures intraventricular wall thickness (mean absolute error [MAE] 1.4mm, 95% CI 1.2-1.5mm), left ventricular diameter (MAE 2.4mm, 95% CI 2.2-2.6mm), and posterior wall thickness (MAE 1.2mm, 95% CI 1.1-1.3mm) and classifies cardiac amyloidosis (area under the curve of 0.83) and hypertrophic cardiomyopathy (AUC 0.98) from other etiologies of LVH. In external datasets from independent domestic and international healthcare systems, EchoNet-LVH accurately quantified ventricular parameters (R2 of 0.96 and 0.90 respectively) and detected cardiac amyloidosis (AUC 0.79) and hypertrophic cardiomyopathy (AUC 0.89) on the domestic external validation site. Leveraging measurements across multiple heart beats, our model can more accurately identify subtle changes in LV geometry and its causal etiologies. Compared to human experts, EchoNet-LVH is fully automated, allowing for reproducible, precise measurements, and lays the foundation for precision diagnosis of cardiac hypertrophy. As a resource to promote further innovation, we also make publicly available a large dataset of 23,212 annotated echocardiogram videos.
2. Department of Medicine, Division of Cardiology, Stanford University
LVH can arise from distinct cardiac and systemic conditions that produce similar appearances, making routine differentiation and precise measurement difficult. EchoNet-LVH is introduced as an end-to-end deep learning approach to quantify ventricular morphology and predict LVH etiology.
- Clinical motivation: Different etiologies, including hypertension, aortic stenosis, hypertrophic cardiomyopathy, and cardiac amyloidosis, can produce similar LVH presentations on routine imaging.Some conditions involve remodeling under physiologic stress, whereas hypertrophic cardiomyopathy and cardiac amyloidosis can increase left ventricular mass without such stress.
- Clinical motivation: Ventricular wall thickness has prognostic value and helps risk-stratify patients for sudden cardiac death and potential defibrillator implantation.
- Clinical motivation: Manual and expert-dependent measurement introduces accuracy limits, test-retest variability, and intra- and inter-provider variability across imaging modalities.These limitations persist even with high-resolution cardiac magnetic resonance imaging, while echocardiography depends on expert interpretation and careful measurement application.
- Rationale: Artificial intelligence-enhanced echocardiography was hypothesized to add value by detecting LVH and predicting its potential etiology.Echocardiography is widely available, low cost, free of ionizing radiation, and foundational to society guidelines for diagnosing hypertrophy.
- Approach: EchoNet-LVH is an end-to-end deep learning workflow that labels left ventricular dimensions, quantifies wall thickness, and predicts LVH etiology.The workflow segments ventricular structures in parasternal long-axis videos, evaluates hypertrophy beat to beat, and uses a three-dimensional residual CNN to predict etiologies including cardiac amyloidosis and aortic stenosis.
Results
EchoNet-LVH uses video-based segmentation and beat-to-beat analysis to automate ventricular measurements, then evaluates those measurements across internal and external healthcare-system datasets. Its measurements showed high agreement with expert annotations and supported higher-fidelity assessment across cardiac cycles.
- Ventricular measurement: EchoNet-LVH applies atrous convolutions to segment ventricular structures in parasternal long-axis echocardiogram videos.Full-resolution PLAX frames are used as input for higher-resolution assessment of LVH.
- Ventricular measurement: Sparse clinical annotations are generalized into predictions for every video frame, enabling beat-to-beat estimation of wall thickness and ventricular dimensions.This approach addresses the common workflow of labeling only one or two frames because annotation is tedious.
- Internal evaluation: R2 0.97 on the held-out SHC test dataset accompanied measurement MAEs of 1.2mm for IVS, 2.4mm for LVID, and 1.4mm for LVPW.These errors were comparable to reported clinical inter-provider variation for the corresponding measurements.
- External evaluation: R2 0.90 on the Unity external test dataset was accompanied by MAEs of 1.6mm for IVS, 3.6mm for LVID, and 2.1mm for LVPW.The external evaluation used 1,791 videos and was performed without model tuning.
- Beat-to-beat assessment: Beat-to-beat assessment can provide higher-fidelity overall measurements by accounting for variation from filling time and irregular heart rate.Automated frame-level measurement also enables rapid, high-throughput analysis that would be tedious with manual tracing.
Prediction of Etiology of Hypertrophy
EchoNet-LVH predicts the etiology of LVH from echocardiogram videos using video-based modeling. It distinguished cardiac amyloidosis, hypertrophic cardiomyopathy, and aortic stenosis from other LVH etiologies in internal testing, with amyloidosis and hypertrophic cardiomyopathy also evaluated externally.
- Cohorts: EchoNet-LVH was trained, validated, and tested on SHC etiology cohorts containing 6,215, 787, and 765 videos, respectively.
- External evaluation: AUC 0.79 for cardiac amyloidosis and AUC 0.89 for hypertrophic cardiomyopathy were obtained on the external CSMC test dataset.The dataset contained 2,351 A4c videos, including 358 amyloidosis, 146 aortic stenosis, 468 hypertrophic cardiomyopathy, and 1,379 other-LVH videos.
Discussion
EchoNet-LVH provides an automated workflow for measuring ventricular structure and predicting LVH etiology from echocardiogram videos. Its performance was robust across external healthcare systems, while broader deployment remains bounded by image-quality and prospective-validation needs.
- EchoNet-LVH provides a fully automated workflow for quantifying left ventricular hypertrophy and predicting its etiology.
- State-of-the-art performance was reported for assessing ventricular thickness and diameter within the variance of human assessment.
- EchoNet-LVH’s multi-beat measurements aid detection of subtle ventricular phenotypes and support clinical test-retest assessment.
- Performance in assessing ventricular thickness was robust across continents, clinical practice patterns, and imaging instrumentation.
- Further work is needed to evaluate performance on more-variable-quality videos acquired with differing expertise and to conduct prospective testing in diverse clinical environments.
- The publicly released dataset contains 22,030 echocardiogram videos with matched human expert annotations for future model comparison and validation.
Data Curation
The study curated echocardiographic videos from routine studies, extracting PLAX and A4C 2D views with expert measurements as labels for ventricular hypertrophy.
- Routine echocardiogram studies contain 50–100 videos and still images spanning multiple cardiac views and acquisition techniques.
- PLAX and A4C 2D videos were extracted from each study for analysis.
- Human expert annotations of IVS, LVID, and LVPW served as training labels for assessing ventricular hypertrophy.
- PLAX videos from SHC were divided into 9,600 training, 1,200 validation, and 1,200 test patients.
- An additional 7,767 SHC studies with defined disease characteristics, including amyloidosis, HCM, aortic stenosis, and hypertension, were identified.
- Videos underwent automated de-identification and labeling, quality checks, view confirmation, and exclusion of color-Doppler videos.
Domestic and International External Health Care System Test Datasets
External evaluation used echocardiographic data from Cedars-Sinai and the Unity Imaging Collaborative, including held-out domestic and international datasets not used during training.
- CSMC and Unity Imaging Collaborative echocardiogram studies were used to evaluate key-point identification and ventricular-dimension measurement.
- A domestic CSMC held-out test dataset contained 3,660 extracted videos.
- The Unity Imaging dataset was a separate international held-out test set not seen during model training.
- Unity echocardiogram videos were obtained from seven British echocardiography laboratories.
EchoNet-LVH Development and Training
EchoNet-LVH combines image-based key-point detection for ventricular measurements with video-based classification of LVH etiology, using temporal processing and beat-level aggregation.
- The models evaluated video length, resolution, and temporal resolution as hyperparameters for performance optimization.
- A modified DeepLabV3 architecture identified ventricular measurement key points in PLAX images by minimizing weighted mean square error.
- An 18-layer ResNet3D classified videos as amyloid or not amyloid using binary cross-entropy loss.
- Test-time augmentation aggregated PLAX predictions across each echocardiogram video.
- Predicted LVID identified peak systole and diastole, enabling systolic and diastolic measurements for every beat.
- Beat-to-beat variation was compared with human-labeled measurements to evaluate method precision.
Comparison with Human Measurement
The evaluation compared EchoNet-LVH measurements with expert annotations and clinician variation across multiple datasets, while assessing beat-to-beat precision and disease-etiology classification.
- Comparison with Human Measurement: Measurements from 23,874 clinically stable paired studies were used to estimate clinician variation between prior and subsequent examinations.
- Comparison with Human Measurement: EchoNet-LVH variation was compared with the variance between repeated clinician measurements.
- Comparison with Human Measurement: The model’s semantic-segmentation measurements of LVID, LVPW, and IVS were evaluated against human labels in a held-out test set.
- Comparison across health systems: Measurement correlations were assessed across SHC, CSMC, and Unity datasets with sample sizes of 1,200, 1,309, and 1,791, respectively.
- Beat-to-beat evaluation: Frame-by-frame predictions automated systole and diastole detection, supporting beat-to-beat ventricular-hypertrophy measurements.
- Beat-to-beat evaluation: Beat-level variation was visualized across 2,320 videos in the internal test dataset.
- Disease Etiology Classification: Disease-classification performance was evaluated with receiver-operating-characteristic and precision-recall curves for amyloidosis, aortic stenosis, and hypertrophic cardiomyopathy.
Extended Figure 1. Hyperparameter Search of Video-based Classification Model
The supplemental material examines video-classification hyperparameters and compares EchoNet-LVH measurements with human annotations across healthcare systems.
- R3D, R2+1D, and MC3 architectures were compared for hypertrophy-etiology classification.
- 8, 16, 32, 64, and 96-frame video clips were evaluated for processing time and memory usage.
- EchoNet-LVH measurements were correlated with human annotations across SHC, CSMC, and Unity healthcare systems.The comparison included n = 1,200 at SHC, n = 1,309 at CSMC, and n = 1,791 at Unity.
- Clinician-reported measurements were compared with prior studies for 23,874 SHC examinations without significant change.
Supplemental Methods
The ventricular-dimension model uses sparse keypoint annotations and a modified loss, then derives measurements from channel centroids while filtering low-quality frames.
- Keypoint prediction: Four diastolic and systolic keypoints are represented as four output channels in a DeepLabV3 architecture.Each video has one labeled frame at diastole and one at systole, with gaussian sampling used to simulate human variation.
- Loss function: A modified mean squared error loss independently penalizes false-positive and false-negative errors using an imbalance-sensitive hyperparameter.Sparse labels contain roughly 4 pixels among 480x640 pixels per frame.
- Loss function: An alpha value of 0.001 and an augmented L2 loss weight of 0.001 were selected through experimentation and hyperparameter sweeps.
- Measurement calculation: Output-channel centroids define measurement endpoints for IVS, LVID, and LVPW calculation.Low-confidence pixels below 0.3 and frames with inconsistent measurement angles greater than 30 degrees are excluded.
- Quality control: Beat-to-beat evaluation uses heuristics to exclude low-quality frames from the overall calculation.