Estimation of Segmental Longitudinal Strain in Transesophageal Echocardiography by Deep Learning

AI summary

Overview

Research area: Medical computer vision / deep learning for echocardiographic motion estimation and cardiac strain quantification.

Technical level: Intermediate. Readers need general familiarity with deep learning, optical flow versus point tracking, and basic echocardiography concepts (strain, left ventricle segments), but the paper's clinical framing is accessible.

Scope: The paper compares two deep learning motion-estimation strategies for estimating segmental longitudinal strain (SLS) of the left ventricle from transesophageal echocardiography (TEE), training and testing them on a newly created synthetic dataset and validating against clinical references from 16 patients.

What This Paper Is About

Segmental longitudinal strain measures how much each individual segment of the left ventricle deforms through the cardiac cycle, and it is a key indicator of regional heart dysfunction such as myocardial ischemia. Today it is measured with speckle tracking echocardiography, which requires substantial manual correction, expertise, and time, and which struggles with the decorrelation and artifacts intrinsic to ultrasound. The authors build and compare the first automated pipeline ("autoStrain") for SLS estimation in TEE, using deep learning motion estimation trained on synthetic data because ground-truth myocardial motion is essentially unavailable for real TEE sequences.

Key Contributions

1. The first comparison of two state-of-the-art deep learning strategies for myocardial motion estimation in TEE: TeeFlow, a RAFT-based dense frame-to-frame optical flow model, and TeeTracker, a CoTracker-based sparse point trajectory model that tracks a myocardial mesh across the full sequence.
2. A novel pipeline (autoStrain) for automatic estimation of regional left ventricular function by predicting SLS in critically ill patients, using the AHA 18-segment model.
3. An open-access database of 240 synthetic TEE sequences from 80 patients with ground-truth myocardial motion, covering four levels of speckle decorrelation and synthetic myocardial infarction.
4. A comprehensive experimental plan, including ablations of pre-trained versus fine-tuned models and single-dataset, sequential, and combined training schemes, to identify the best-performing approach on real data.

Main Findings

• TeeTracker outperforms TeeFlow: On the synTEE test data, TeeTracker achieved a mean distance error of 0.65 ± 0.20 mm, compared with 1.55 ± 0.56 mm for the best TeeFlow model (average across test sets). Even the pre-trained TeeTracker (1.08 mm) surpassed the best fine-tuned TeeFlow (1.55 mm), which the authors state is the first clear demonstration of point trajectory estimation beating dense motion estimation in echocardiography.
• Fine-tuning on synthetic TEE matters: Mean distance error dropped from 3.28 mm to 1.55 mm for TeeFlow and from 1.08 mm to 0.65 mm for TeeTracker after fine-tuning on the simulated sequences.
• Decorrelation drives difficulty: TeeTracker fine-tuned on the least-decorrelated dataset reached 0.36 ± 0.06 mm on test set 1 but degraded to 1.60 mm on the most-decorrelated test set 4. Training on the most decorrelated data improved test set 4 to 1.29 ± 0.41 mm but worsened performance on easier data (0.73 mm on test set 1).
• Combined training was best: Training on all four synthetic datasets simultaneously gave the best overall average error (0.65 ± 0.20 mm), better than sequential training (0.97 ± 0.33 mm).
• Synthetic infarction improves detection: Adding simulated infarction data reduced mean distance error to 0.37 ± 0.06 mm across all segments and 0.36 ± 0.11 mm on infarcted segments, versus 0.58 ± 0.14 mm without it. SLS agreement improved to a mean difference of 0.14% (limits of agreement -2.60% to 2.88%) and GLS to 0.11% (-0.69% to 0.91%).
• Synthetic strain agreement: Across all synTEE test data, autoStrain reached a GLS mean difference of 2.78% (-1.63% to 7.19%) and an SLS mean difference of -0.38% (-5.00% to 4.25%). Restricting to basal and mid segments improved this to -0.22% (-3.53% to 3.09%).
• Clinical validation: On 16 patients, TeeTracker trained with synthetic infarcts gave the best SLS agreement across all segments, with a mean difference (95% limits of agreement) of 1.09% (-8.90% to 11.09%), and GLS of -2.36% (-8.36% to 3.52%).
• Apical segments are harder: Excluding apical segments gave more contrasted results: TeeFlow achieved the best mean difference of -0.14%, while TeeTracker had the lowest standard deviation at 5.10%.
• Speed: TeeFlow ran at a mean 5.60 frames per second and TeeTracker at 1.59 frames per second, both far faster than manual analysis, which the authors state typically requires 1-5 minutes per cardiac cycle.

Methodology in Plain English

The team acquired 2D TEE images from 80 patients with a passively placed, unlocked probe, producing the foreshortening, noise, and out-of-plane motion typical of perioperative monitoring. Because real TEE sequences with ground-truth motion references do not exist, they built synthetic data instead: they estimated scatterer maps from real TEE frames, fed those into a physical ultrasound simulator (SIMUS), and beamformed the raw output into realistic synthetic B-mode sequences with known myocardial contraction fields.

For each patient they generated four sequences sharing the same motion but with different speckle patterns, by varying the ratio of coherent to incoherent scatterers (0.9, 0.7, 0.6, and 0.5), producing datasets from mildly to severely decorrelated. They also simulated myocardial infarction by reducing longitudinal contraction in a localized segment following a Gaussian distribution, with surrounding tissue compensating to preserve overall contraction.

Two models were trained on this data. TeeFlow adapts RAFT, estimating dense displacement fields between consecutive frames. TeeTracker adapts CoTracker, using a transformer to track a sparse set of mesh points over a sliding window of 8 frames, starting from the end-systole frame and tracking both backward and forward. Both were fine-tuned from pre-trained weights and evaluated by mean Euclidean distance between estimated and reference meshes, then plugged into the autoStrain pipeline, which converts mesh motion into per-segment longitudinal strain as a percentage of end-diastolic length.

Why This Matters

The paper argues that automated SLS from TEE could make regional cardiac function assessment objective, reproducible, and practical in settings where it currently is not. Strain analysis today depends on expert-driven speckle tracking with manual corrections, making it too resource-intensive for continuous monitoring.

Real-world applications:

• Perioperative monitoring of mechanically ventilated patients, where TEE is preferable to transthoracic imaging and the probe can remain passive in the esophagus for continuous surveillance.
• Early detection of myocardial ischemia and regional wall motion abnormalities in interventional cardiology, electrophysiology, and cardiothoracic surgery.
• Objective, repeatable quantification of segmental function that does not depend on individual operator expertise or manual boundary adjustments.
• Potential triage or screening support in intensive care, where strain estimates could flag hypokinetic segments for specialist review.

Industry relevance: the work targets the same clinical territory as commercial speckle tracking packages (the paper uses GE EchoPAC for reference annotations and a GE Vivid E95 scanner with a 6VT-D probe), and its release of an open synthetic TEE dataset gives ultrasound vendors and imaging AI developers a benchmark for TEE-specific motion estimation, a domain where synthetic datasets previously existed mainly for transthoracic imaging.

Future Directions

• Closing the gap on apical segments: the best SLS agreement came only after excluding the apex, so improving apical tracking remains an open problem.
• Extending clinical validation beyond the 16 patients reported, and testing whether the synthetic-infarction training benefit holds in a wider range of real pathologies.
• Improving inference speed: TeeTracker's 1.59 frames per second may limit real-time deployment, even though it is far faster than manual analysis.
• Broadening the simulation to cover other pathologies and more complex deformation patterns; the reported synthetic infarction reduces contractility in one of six cardiac segments following a Gaussian profile, which is a simplification of real disease.
• The provided paper content is truncated mid-discussion, so the authors' own stated limitations and future work are not fully reported here.

Target Audience

This paper is most useful to medical imaging and computer vision researchers working on motion estimation, tracking, or ultrasound analysis; cardiologists and cardiac anesthesiologists interested in perioperative strain monitoring; and engineers at ultrasound vendors or medical AI companies building automated strain quantification tools. Readers seeking a rigorous clinical trial of strain software will not find it here, since the clinical validation covers 16 patients and the primary evaluation is on synthetic data.

Authors’ abstract

Segmental longitudinal strain (SLS) of the left ventricle (LV) is an important prognostic indicator for evaluating regional LV dysfunction, in particular for diagnosing and managing myocardial ischemia. Current techniques for strain estimation require significant manual intervention and expertise, limiting their efficiency and making them too resource-intensive for monitoring purposes. This study introduces the first automated pipeline, autoStrain, for SLS estimation in transesophageal echocardiography (TEE) using deep learning (DL) methods for motion estimation. We present a comparative analysis of two DL approaches: TeeFlow, based on the RAFT optical flow model for dense frame-to-frame predictions, and TeeTracker, based on the CoTracker point trajectory model for sparse long-sequence predictions. As ground truth motion data from real echocardiographic sequences are hardly accessible, we took advantage of a unique simulation pipeline (SIMUS) to generate a highly realistic synthetic TEE (synTEE) dataset of 80 patients with ground truth myocardial motion to train and evaluate both models. Our evaluation shows that TeeTracker outperforms TeeFlow in accuracy, achieving a mean distance error in motion estimation of 0.65 mm on a synTEE test dataset. Clinical validation on 16 patients further demonstrated that SLS estimation with our autoStrain pipeline aligned with clinical references, achieving a mean difference (95\% limits of agreement) of 1.09% (-8.90% to 11.09%). Incorporation of simulated ischemia in the synTEE data improved the accuracy of the models in quantifying abnormal deformation. Our findings indicate that integrating AI-driven motion estimation with TEE can significantly enhance the precision and efficiency of cardiac function assessment in clinical settings.