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, et al.
eess.IV · cs.AI · cs.CV
Quick verdict
TAS-Net-D adds persistent tokens and drift-aware fine-tuning to reduce strain drift in echo tracking; promising, but small cohort and no significance tests.
Summary
This method paper targets temporal drift in deep-learning myocardial tracking for echocardiography. Deep trackers adapted from optical flow lack cyclic constraints, so tracked points may not return to their initial relative positions at end-diastole, biasing strain estimates. The authors extend TAS-Net (a CoTracker1-based tracker) into TAS-Net-D with two architectural changes: refinement iterations interleaved across sliding windows, and window-persistent tokens carrying information across windows as the sole gradient pathway between them, at computation matched to the baseline. A teacher-student fine-tuning on real CAMUS/CARDINAL videos combines a drift loss (edge-length consistency between end-diastolic frames, stop-gradient normalized), visibility-weighted point-to-point and Chamfer anchor losses to the frozen teacher, and a third-derivative smoothness regularizer. On the HUNT test-retest dataset (30 patients), TAS-Net-D achieves the lowest mean trajectory error (1.22±0.01 mm vs 1.28 for TAS-Net), lower GLS drift (0.98±0.06% vs 1.20±0.05%), and the best test-retest agreement (MAE 1.83%, σ 2.33), slightly better than the clinical reference itself (1.99, 2.51). Multi-cycle Bland-Altman on 120 CAMUS videos shows the lowest cycle-to-cycle variance. The authors acknowledge EchoTrackerR, trained on EchoPac-derived labels, retains better reference GLS agreement; no statistical significance tests are reported, and ablated variants outperform the full model on some test-retest metrics.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Window-persistent tokens plus interleaved cross-window refinement let tracking information flow over the full cardiac cycle at computation matched to TAS-Net (Sect. 2.1).
- Teacher-student fine-tuning on real echocardiographic videos, without motion labels, enforces cyclic consistency via a drift loss plus anchor and smoothness terms (Sect. 2.2).
- On HUNT, TAS-Net-D reduces MTE (1.22±0.01 vs 1.28±0.00 mm) and GLS drift (0.98±0.06% vs 1.20±0.05%) relative to TAS-Net, and achieves the best test-retest reproducibility (MAE 1.83%, σ 2.33), slightly better than the clinical reference (1.99, 2.51).
- A posteriori linear drift correction zeroes drift by construction but degrades MTE and does not improve agreement or reproducibility, motivating in-the-loop correction.
- In multi-cycle CAMUS evaluation, TAS-Net-D shows the lowest cycle-to-cycle GLS/RLS variance (fig. 3).
Strengths
Real clinical problem with a precise failure criterion
Sect. 1 frames drift as strain failing to return to zero at end-diastole (ED), the exact quantity clinicians read; Table 1 even reports the reference's own drift floor (0.32) so improvements are interpretable.
Computation-matched, mechanism-transparent architecture change
Sect. 2.1 sets Ii=2, Io=2 (train)/3 (inference) to preserve TAS-Net's total refinement budget, and persistent tokens are detached only at outer-iteration boundaries, making them the sole cross-window gradient pathway — the comparison isolates information flow, not extra compute.
Carefully designed drift loss
Honest baseline framing including an unfavorable result
Clinically relevant evaluation suite
Weaknesses
No statistical inference anywhere
The paper's own ablations undercut the loss-design claim
Reference circularity in evaluation
No architecture-only ablation
Dependence on ED/ES instants is unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.09577, MICCAI-style) addresses temporal drift in deep-learning myocardial tracking for echocardiography: without cyclic constraints, tracked points fail to return to their initial relative positions at end-diastole (ED), corrupting global and regional longitudinal strain (GLS/RLS). Central claims: (1) extending TAS-Net with window-persistent tokens and interleaved cross-window refinement enables full-cycle information flow at computation matched to the baseline (Sect. 2.1); (2) teacher-student fine-tuning on real videos with a drift loss enforces cyclic consistency without motion labels (Sect. 2.2); (3) on HUNT test-retest, TAS-Net-D reduces drift and MTE and achieves the best reproducibility, including vs the clinical reference (Table 1); (4) multi-cycle coherence improves on 120 CAMUS videos (figs. 3-4). Code is released.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is temporal drift in deep learning myocardial tracking?
Because deep trackers adapted from optical flow lack a cyclic constraint, small per-frame errors accumulate over the cardiac cycle: tracked points may not return to their initial relative positions at end-diastole, so the computed strain does not return to zero and can even diverge from cycle to cycle.
Does TAS-Net-D beat commercial speckle tracking software like EchoPac?
It is evaluated against EchoPac-derived reference tracings on the HUNT dataset. Its test-retest reproducibility (MAE 1.83%, σ 2.33) was slightly better than the reference's own (1.99, 2.51), but its agreement with the EchoPac reference (GLS MAE 1.81) was worse than EchoTrackerR (1.53), and no statistical significance tests were reported.
Is the code or data for this paper available?
Code is released at github.com/ThierryJudge/cardiac-motion. Training uses real CAMUS/CARDINAL videos plus the simulated TAS-1K set; the HUNT test-retest evaluation data (30 patients) is clinical and not publicly available.
Why is a posteriori drift correction not enough?
The paper's linear drift correction drives drift to zero by construction, but it increases mean trajectory error (1.34 vs 1.28 mm for plain TAS-Net) and gives no reproducibility gain (test-retest MAE 1.89 vs 1.83), because it redistributes existing tracking error instead of fixing the tracking itself.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table 1's column headers interleave during PDF extraction; the metric-to-column assignment cited here (MTE / GLS drift / RLS drift / reference agreement / test-retest) was cross-checked against five quantitative statements in the prose, but mid-table μ/σ groupings may be misaligned.
- Figures are present only as captions in the extraction; the Bland-Altman spread (fig. 3) and strain maps (fig. 4) could not be visually inspected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.