Source-linked AI summary
Myocardial Strain Drift Correction in Deep Learning Based Ultrasound Tracking
Thierry Judge, Nicolas Duchateau, Andreas Østvik, Havard Dalen, Bjørnar Grenne, Pierre-Yves Courand, Lasse Lovstakken, Pierre-Marc Jodoin, Olivier Bernard
TL;DR
Deep learning myocardial trackers often lack physiological constraints, causing drift that prevents tracked points and strain from returning consistently to their initial end-diastolic state. The paper extends TAS-Net with window-persistent tokens and teacher-student fine-tuning to enforce cyclic motion, reducing drift and improving test–retest robustness. The authors conclude that this advances more reliable automatic strain estimation while retaining scope for future segmentation-based initialization.
Problem
Deep learning myocardial trackers often lack physiological constraints, so accumulated sequence errors can produce nonzero or divergent strain across cardiac cycles.
Method
The framework extends TAS-Net with window-persistent tokens and teacher-student fine-tuning using drift, shape-preserving, and temporal regularization losses.
Results
TAS-Net-D reduces drift and improves robustness in test–retest and multi-cycle evaluations.
Takeaways & Limitations
The method is a step toward fully automatic strain estimation, with automatic segmentation identified as future work to remove manual initialization.
Takeaways & Limitations
Zero drift may not always be desirable because acquisition variability can induce apparent geometric changes, and out-of-plane rotation can affect projected edge distances.
Abstract
from arXiv · showhide
Myocardial strain from echocardiography is a key biomarker for cardiac function. Recent deep learning methods show strong performance for myocardial motion tracking but often lack physiological constraints, leading to temporal drift across the cardiac cycle. Consequently, tracked points may not return to their relative initial positions at the end of each cardiac cycle, producing inaccurate strain estimates and even divergence in some cases. We propose a deep learning framework that compensates for drift during myocardial tracking. We extend a state-of-the-art echocardiographic tracking method (TAS-Net) with persistent memory tokens that share information across sliding windows over full cardiac cycles. A teacher-student fine-tuning strategy on real echocardiographic data then enforces physiologically consistent cyclic motion while preserving tracking accuracy. Experiments show reduced global and regional strain drift, improved agreement with clinical references, and better test-retest reproducibility, supporting more reliable myocardial strain estimation in clinical practice.
1 Introduction
Myocardial strain is informative for assessing cardiac dysfunction, but existing echocardiographic software and deep learning trackers have important limitations. The proposed framework addresses temporal drift by enforcing physiologically consistent cyclic motion during learning.
- Myocardial strain quantifies cardiac-wall deformation across the cycle and informs assessment of cardiac dysfunction.
- Clinical speckle-tracking software lacks transparent, standardized post-processing, is not freely available, and may perform poorly on lower-quality images.
- Deep learning trackers can achieve competitive results but often lack physiological constraints, allowing errors to accumulate across cycles as strain drift.
- The proposed framework extends TAS-Net with long-range temporal mechanisms and teacher-student fine-tuning on real echocardiographic data to reduce drift without compromising tracking accuracy.
2 Method
TAS-Net-D extends sliding-window myocardial tracking with cross-window information flow and teacher-student losses designed to reduce cyclic drift while preserving tracking consistency.
- Model architecture: TAS-Net-D addresses limited temporal context by interleaving refinement across windows and adding persistent tokens that carry information through the full video.Refinement is distributed across windows, while persistent tokens provide the cross-window information pathway.
- Model architecture: Persistent tokens are updated from current appearance features and read through cross-attention, allowing information to accumulate across sliding windows.They are detached only at the end of each outer iteration, making them the sole gradient pathway across windows.
- Training: The teacher-student training setup uses a pretrained TAS-Net teacher and real videos because simulated motion can contain compounded drift.The CNN encoder is frozen to reduce memory use and permit full-video processing.
- Losses: The drift loss compares neighboring-mesh edge lengths between two end-diastolic instants, targeting cyclic consistency beyond longitudinal strain alone.The mesh edge set includes neighboring radial and longitudinal nodes.
- Losses: Anchor losses preserve teacher-consistent global contour shape and point locations, with pointwise supervision weighted by teacher visibility.Visibility weighting emphasizes reliable points while allowing drift correction to dominate in low-visibility regions.
- Losses: A temporal regularization loss penalizes rapid changes in the third derivative of edge length, preventing drift correction from producing strain-curve spikes.The regularizer complements the drift loss, which acts only at the two end-diastolic instants.
3 Experimental setup
The evaluation compares myocardial tracking methods on simulated and real echocardiographic data, including drift correction and test–retest performance against EchoPac references.
- Experimental setup: Five tracking approaches and an a posteriori TAS-Net correction were compared, including models trained on simulated or real echocardiographic videos.The simulated training set contained 1,478 videos, while EchoTrackerR used 6,490 real videos.
- Drift correction: The linear correction was anchored at end-systole and distributed drift symmetrically across the cardiac cycle while preserving the initialization mesh.
- TAS-Net-D training: TAS-Net-D was trained on corresponding real CAMUS and CARDINAL videos for 100,000 AdamW steps with a one-cycle learning-rate schedule.The initial learning rate was 1 × 10−4, with heuristically selected loss weights.
- Evaluation data: Evaluation used 30 HUNT patients with two independent acquisitions in each of three apical views and clinical expert annotations from EchoPac.The design enabled assessment of inter-expert variability and test–retest reproducibility.
- Evaluation metrics: Methods were evaluated using mean trajectory error, peak GLS mean absolute error, Bland–Altman agreement, and average test–retest variability.
4 Results
On HUNT, full-video trackers showed low drift but weaker whole-cycle alignment or reproducibility, whereas TAS-Net-D improved agreement and test–retest performance without relying on post hoc correction. Multi-cycle CAMUS examples further show more coherent strain behavior for TAS-Net-D.
- HUNT results: Full-video MyoTracker and CoTracker3 achieved low GLS and RLS drift and good GLS agreement, but showed poorer whole-cycle alignment and lower test–retest reproducibility.GLS compares only two frames, so it may not represent alignment over the entire cycle.
- Multi-cycle results: In CAMUS A4C multi-cycle Bland–Altman plots, TAS-Net-D had a mean closer to zero and smaller variance for GLS and RLS.
- HUNT results: TAS-Net-D slightly improved GLS agreement over TAS-Net, had lower MTE than other methods, and achieved the best test–retest MAE and variance.Its reference agreement remained poorer than EchoTrackerR, which was trained on EchoPac-derived labels.
- Drift correction comparison: Post hoc linear correction reduced GLS and RLS drift to zero by construction, but increased MTE without improving reference agreement or test–retest reproducibility.
- Multi-cycle results: TAS-Net-D produced more coherent two-cycle GLS and RLS curves and strain maps, with greater spatial consistency in the maps.Spatial consistency was not explicitly optimized.
5 Conclusion
TAS-Net-D extends TAS-Net to reduce drift by passing information across sliding windows and fine-tuning on real echocardiographic videos. It improves robustness across test–retest and multi-cycle evaluations, while remaining sensitive to acquisition-related geometric variation.
- Conclusion: TAS-Net-D mitigates drift by extending TAS-Net beyond its sliding-window formulation for deep learning myocardial tracking.
- Conclusion: Modified refinement and window-persistent tokens let information flow across windows during fine-tuning on real echocardiographic videos.
- Conclusion: Multi-cycle CAMUS examples showed greater coherence between cycles for strain curves and maps, plus increased spatial consistency in strain maps.The spatial consistency improvement was not explicitly optimized.
- Conclusion: TAS-Net-D reduced drift and improved robustness in test–retest and multi-cycle evaluations.
- Limitations and scope: Zero drift may not be desirable because acquisition variability can produce apparent geometric changes; out-of-plane rotation can still affect projected edge distances.The edge-length metric is invariant to in-plane rotation and translation.