EchoTracker: Myocardial Point Tracking

• EchoTracker is a learning-based myocardial point tracking method that formulates tracking as a sparse, full-sequence problem rather than dense frame-to-frame optical flow.
• It utilizes a two-fold coarse-to-fine architecture, achieving an average position accuracy of 67% and a median trajectory error of 2.86 pixels.
• The system boosts global longitudinal strain reproducibility by approximately 25% while efficiently processing videos in 0.24 seconds per video on modern GPUs.

EchoTracker is a learning-based myocardial point tracking method for echocardiography that tracks queried tissue points on the left ventricular myocardium through an ultrasound cine loop. It formulates myocardial motion estimation as a sparse point-tracking problem over a full image sequence rather than as dense frame-to-frame optical flow, and is organized as a “two-fold coarse-to-fine model” with a coarse trajectory initializer followed by iterative fine reinforcement. In the original report, EchoTracker achieved an average position accuracy of $67\%$, a median trajectory error of $2.86$ pixels, and a relative improvement of about $25\%$ for global longitudinal strain (GLS) reproducibility in a clinical test-retest setting (Azad et al., 2024).

1. Formal task and problem setting

EchoTracker is defined on an ultrasound sequence

$$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$

and a set of query points in the first frame

$p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$

The required output is the set of tracked positions across all frames,

$P = \{p_s \in (x_s^n,y_s^n)\},$

summarized in the paper as

$\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$

This is a sparse point-tracking formulation rather than dense optical flow estimation, with the points assumed to lie on myocardial tissue surfaces and then propagated through the cardiac cycle (Azad et al., 2024).

The formulation is clinically motivated by deformation analysis, especially strain imaging such as GLS. The paper explicitly rejects the occlusion semantics used in TAP-Vid, where point location becomes meaningless under occlusion, because echocardiographic strain computation still requires a location estimate even when tissue moves partially out of plane. EchoTracker therefore follows the “persistent particles” viewpoint associated with PIPs, in which points are always tracked. This makes the method structurally aligned with myocardial strain measurement, where coordinate continuity is operationally more important than binary visibility labeling.

2. Why myocardial tracking in ultrasound is difficult

The underlying difficulty arises from the interaction of cardiac biomechanics and ultrasound image formation. The myocardium contracts, relaxes, translates, and deforms cyclically, so the motion is non-rigid and complex. At the same time, ultrasound appearance is degraded by speckle, noise, low contrast, decorrelation, and acoustic artifacts. Points may also become weakly visible or effectively disappear because of occlusion, out-of-plane motion, or changing insonation. Over a full cardiac cycle, these effects accumulate into drift (Azad et al., 2024).

The paper positions EchoTracker against two families of prior methods. First, conventional optical flow methods, including FlowNet, PWC-Net, and echocardiography-specific variants, estimate dense displacement mainly between consecutive frames. Because they do not explicitly model long temporal context, errors compound over time, especially under speckle decorrelation and out-of-plane motion. Second, modern point trackers from computer vision—TAP-Net, PIPs, TAPIR, PIPs++, CoTracker, and OmniMotion—offer stronger long-range correspondence modeling, but were not designed specifically for echocardiographic ultrasound. The paper argues that sliding temporal windows, computational growth with longer context, and insufficient adaptation to ultrasound motion statistics limit their direct applicability.

According to the authors, EchoTracker was the first application of “tracking-any-point” ideas in medical imaging (Azad et al., 2024). A plausible implication is that the method represents a domain adaptation of general point-tracking concepts to a regime in which temporal persistence, local deformability, and clinically usable coordinate trajectories are all required simultaneously.

3. Two-fold coarse-to-fine architecture

EchoTracker is organized into two stages. The first stage performs coarse initialization of trajectories from low-resolution, sequence-level evidence. The second stage performs iterative reinforcement from higher-resolution local appearance changes. The “coarse-to-fine” description refers both to spatial resolution and to tracking strategy: a global coarse search reduces ambiguity, then a local fine search corrects the trajectory (Azad et al., 2024).

In addition to the ultrasound images, the model uses “frame flow,” defined as

$u_s-u_{s-1},$

which is raw frame differencing rather than optical flow. In the initialization stage, each frame is encoded by a pruned 2D residual convolutional encoder, producing coarse feature maps

$$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$

with downsampling factor $k=8$ and feature dimension $2.86$0. For each query point $2.86$1, a query descriptor is obtained by bilinear interpolation,

$2.86$2

The descriptor is correlated against multiscale frame features through the cost volume

$2.86$3

where the pyramid has $2.86$4 levels and kernel size $2.86$5. A coarse 1D ResNet then operates across time to estimate an initial trajectory over the full sequence. The paper reports that this initialization stage alone achieved

$2.86$6

on a limited subset of DS-B, already surpassing baseline PIPs and TAP-Net.

The fine stage reuses the same basic encoder with lighter downsampling, $2.86$7, giving

$2.86$8

Instead of global matching to the frame-0 template, EchoTracker performs local matching around the current trajectory estimate: $2.86$9 To improve robustness to appearance change, additional cost volumes from fixed temporal offsets and the first frame are included: $25\%$0 These cost volumes are concatenated, mapped to score maps, and processed by a deeper fine 1D ResNet that predicts trajectory corrections

$25\%$1

The update rule is iterative: $25\%$2 with $25\%$3 iterations during both training and evaluation. The final architecture does not use attention, explicit warping, or transformers. The paper states that replacing the fine ResNet with a transformer reduced performance.

4. Supervision, datasets, and implementation

EchoTracker was trained on real ultrasound trajectories rather than synthetic data. The supervision procedure was semi-supervised: a traditional tracking algorithm first produced candidate tracks, then clinical experts tuned them for optimal tracking and removed points that did not properly follow tissue. The resulting datasets focus on left ventricular myocardium in apical four-chamber, two-chamber, and long-axis views, all acquired on GE Vivid E95 scanners (Azad et al., 2024).

Split	Patients / videos	Notes
DS-A	40 / 210	Independent test-retest test set
DS-B	913 / 2731	Training subset
DS-C	615 / 1837	Training subset
DS-D	643 / 1922	Training subset

DS-A contains, on average, 73 tracked points per video, 93 frames per video, image height 593 pixels, and width 571 pixels, with the two scans for each patient acquired in immediate succession by different operators. DS-B has 81 points per video on average and 85 frames per video. DS-C has 80 points per video on average and 84 frames per video. DS-D has 66 points per video on average and 84 frames per video. The test-retest structure of DS-A is clinically relevant because physiology is expected to remain nearly unchanged while acquisition variability remains measurable.

All images were resized to $25\%$4 for training and evaluation. Training used all available points in each video, batch size 1, AdamW with learning rate

$25\%$5

and a one-cycle learning-rate scheduler. The staged curriculum was DS-B for 22 epochs with sequence length $25\%$6, then DS-C for 50 epochs with $25\%$7, then DS-D for 1 epoch with $25\%$8. Implementation was in PyTorch on an NVIDIA GeForce RTX 3090 with 24 GB memory. The paper states that 50 epochs on DS-B typically take over one week on a single GPU.

The training objective is only partially specified in the paper. It states that all refinement iterations are supervised end to end using the same loss function as in PIPs++, but does not reproduce the exact loss equation or loss weights. No additional preprocessing, augmentation, or appendix-level training equations are described in the provided text.

5. Quantitative evaluation and clinical strain analysis

Evaluation followed point-tracking metrics from TAP-Vid and PointOdyssey. The principal tracking metric is average position accuracy,

$25\%$9

computed over five pixel thresholds $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$0. Individual accuracies

$$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$1

were also reported, together with median trajectory error (MTE) and average inference time (AIT). For clinical assessment, the paper evaluated peak GLS, defined in words as the relative change of longitudinal ventricular length from end-diastole to minimum ventricular length (Azad et al., 2024).

Without fine-tuning on echocardiographic data, the benchmarked point trackers on DS-A achieved $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$2 average position accuracy for PIPs, $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$3 for TAP-Net, $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$4 for TAPIR, $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$5 for PIPs++, and $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$6 for CoTracker, with corresponding MTE values of $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$7, $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$8, $$U = \{u_s \in \mathbb{R}^{H \times W}\}, \quad s=0,1,\dots,S,$$9, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$0, and $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$1 pixels. After fine-tuning, EchoTracker was best across every reported tracking metric on DS-A.

Method	Tracking metrics	Speed
PIPs++	$p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$2, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$3, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$4, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$5, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$6, $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$7, MTE $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$8 px	AIT $p_0=\{(x_0^n,y_0^n)\}, \quad n=0,1,\dots,N.$9 s/video
CoTracker	$P = \{p_s \in (x_s^n,y_s^n)\},$0, average $P = \{p_s \in (x_s^n,y_s^n)\},$1, MTE $P = \{p_s \in (x_s^n,y_s^n)\},$2 px	AIT $P = \{p_s \in (x_s^n,y_s^n)\},$3 s/video
EchoTracker	$P = \{p_s \in (x_s^n,y_s^n)\},$4, average $P = \{p_s \in (x_s^n,y_s^n)\},$5, MTE $P = \{p_s \in (x_s^n,y_s^n)\},$6 px	AIT $P = \{p_s \in (x_s^n,y_s^n)\},$7 s/video

The paper’s “average position accuracy of 67%” therefore means that 67% of all point-frame predictions fall within the specified thresholds when the five-threshold accuracies are averaged. Relative to the strongest fine-tuned baseline, PIPs++, EchoTracker improves average position accuracy from $P = \{p_s \in (x_s^n,y_s^n)\},$8 to $P = \{p_s \in (x_s^n,y_s^n)\},$9, reduces MTE from $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$0 to $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$1 pixels, and reduces AIT from $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$2 to $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$3 seconds per video.

Clinical evaluation focused on GLS agreement and reproducibility. For reference comparison, EchoTracker achieved bias $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$4, standard deviation $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$5, and mean absolute deviation $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$6. In the test-retest setting it achieved $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$7, $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$8, and $\text{EchoTracker}(u_s,p_0)=p_s \in (x_s^n,y_s^n).$9. PIPs++ gave test-retest $u_s-u_{s-1},$0 and $u_s-u_{s-1},$1, while CoTracker gave $u_s-u_{s-1},$2 and $u_s-u_{s-1},$3. EchoPWCNet, included for context, reported test-retest $u_s-u_{s-1},$4 and $u_s-u_{s-1},$5. The abstract’s claim of a $u_s-u_{s-1},$6 relative improvement in GLS estimation is numerically consistent with the reduction from $u_s-u_{s-1},$7 to $u_s-u_{s-1},$8, which is approximately $u_s-u_{s-1},$9 and rounds to $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$0.

6. Strengths, limitations, and clinical significance

The strengths claimed for EchoTracker are explicit. It addresses long-range temporal tracking rather than only consecutive-frame motion; it is adapted to ultrasound rather than transferred unmodified from natural-video tracking; it combines strong coarse initialization with efficient local refinement; it uses a lightweight architecture based on pruned 2D encoders and 1D temporal ResNets; it can operate on arbitrary sequence lengths and numbers of query points subject to memory; it improves both technical tracking metrics and clinically relevant GLS reproducibility; and its reported inference time of $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$1 seconds per video on an RTX 3090 is compatible with practical deployment (Azad et al., 2024).

The limitations are also stated directly. High-quality annotated ultrasound trajectories are scarce, which constrains scale and diversity. All reported data come from GE Vivid E95 scanners and specific apical views, so cross-vendor robustness is not established. Performance likely remains dependent on image quality because ultrasound artifacts and poor visibility are intrinsic. Models trained on shorter sequences perform poorly on longer ones, indicating sensitivity to temporal training distribution. CoTracker’s degraded fine-tuning result suggests unresolved preprocessing or implementation effects. The paper does not provide a full failure-mode analysis, and it does not explicitly solve out-of-plane motion; instead, it enforces persistent point localization because GLS requires coordinates even under compromised visibility.

Clinically, the significance is not merely that trajectories are closer to reference annotations. The paper’s central implication is that improved temporal consistency in myocardial point trajectories yields more stable deformation curves and more reproducible peak GLS. This suggests a route toward more reliable follow-up studies and treatment monitoring, since strain variability becomes less entangled with operator and acquisition variability. That implication is explicitly framed as potential improvement in the diagnostic and prognostic value of echocardiographic functional measurements.

7. Subsequent comparative developments

Later work treated EchoTracker as a domain-specific baseline and, in some cases, as a design point to be simplified or revised. “Low Complexity Point Tracking of the Myocardium in 2D Echocardiography” described EchoTracker as a specialized echocardiographic point tracking architecture originally validated for left-ventricular tracking, noted that it processes the whole sequence at once, uses frame flow and point locations in the update module, and retains four refinement iterations. In that study, the trained EchoTracker model had $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$2M parameters, a $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$3 MB file size, $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$4 MB GPU memory use, and $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$5 ms inference time on a 100-frame, 100-point benchmark, while underperforming the proposed MyoTracker on right-ventricular tracking and RV free wall strain estimation (Chernyshov et al., 13 Mar 2025).

“Taming Modern Point Tracking for Speckle Tracking Echocardiography via Impartial Motion” revisited EchoTracker from the perspective of training bias rather than architecture alone. On dataset D, baseline EchoTracker improved from $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$6 and MTE $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$7 px to $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$8 and $$F_s \in \mathbb{R}^{d \times \frac{H}{k} \times \frac{W}{k}},$$9 px after re-fine-tuning; on OOD dataset E it improved from $k=8$0 and $k=8$1 px to $k=8$2 and $k=8$3 px. The abstract of that paper further reports that EchoTracker “boosts overall position accuracy by 60.7% and reduces median trajectory error by 61.5% across heart cycle phases,” indicating that temporal initialization and directional motion bias were major factors in robustness (Azad et al., 14 Jul 2025).

“EchoTracker2: Enhancing Myocardial Point Tracking by Modeling Local Motion” then questioned whether the coarse initialization stage was necessary at all in myocardial point tracking. In its ablation, EchoTracker with initialization versus without initialization changed only marginally: $k=8$4 from $k=8$5 to $k=8$6, $k=8$7 from $k=8$8 to $k=8$9, $2.86$00 from $2.86$01 to $2.86$02, MTE from $2.86$03 to $2.86$04, and AIT from $2.86$05 to $2.86$06 seconds. EchoTracker2 used that result to motivate a fine-stage-only redesign and reported average gains of $2.86$07 in position accuracy and $2.86$08 in MTE relative to EchoTracker across its evaluation sets (Azad et al., 12 May 2026).

Taken together, these follow-on studies place EchoTracker in a specific historical role. It established full-sequence, ultrasound-specific myocardial point tracking as a practical alternative to both frame-to-frame optical flow and generic point trackers, but subsequent work suggests that its main lasting contribution may be less the exact two-stage architecture than the broader reframing of speckle tracking echocardiography as long-range sparse point tracking with clinically constrained persistence.