Low Complexity Point Tracking of the Myocardium in 2D Echocardiography

TL;DR

This work addresses the need for fast, resource-efficient myocardium tracking in 2D echocardiography by introducing MyoTracker, a low-parameter (approximately $0.32$M) point-tracking network derived from CoTracker2 with extended temporal context and simplified components. It processes entire sequences in one pass, drastically reducing memory usage and latency while maintaining high tracking accuracy for the RV free wall and providing RV FWS estimates with bias $$-0.3\%$$ and 95% limits of agreement $[-6.1\%, 5.4\%]$ that are within interobserver variability. Compared to CoTracker2 and EchoTracker, MyoTracker achieves lower trajectory errors and higher inference speed, requiring about $67\%$ less GPU memory than CoTracker2 and $84\%$ less than EchoTracker for long sequences, and it is $74\times$ faster than CoTracker2 and $11\times$ faster than EchoTracker in the authors’ setup. An ablation study shows the most significant accuracy gains arise from enabling full-sequence temporal context, with trade-offs between refinement iterations and runtime. The results support deploying MyoTracker in clinical workflows and mobile devices to accelerate RV measurements and enable parallelized processing, while maintaining clinically acceptable strain estimations.

Abstract

Deep learning methods for point tracking are applicable in 2D echocardiography, but do not yet take advantage of domain specifics that enable extremely fast and efficient configurations. We developed MyoTracker, a low-complexity architecture (0.3M parameters) for point tracking in echocardiography. It builds on the CoTracker2 architecture by simplifying its components and extending the temporal context to provide point predictions for the entire sequence in a single step. We applied MyoTracker to the right ventricular (RV) myocardium in RV-focused recordings and compared the results with those of CoTracker2 and EchoTracker, another specialized point tracking architecture for echocardiography. MyoTracker achieved the lowest average point trajectory error at 2.00 $\pm$ 0.53 mm. Calculating RV Free Wall Strain (RV FWS) using MyoTracker's point predictions resulted in a -0.3$\%$ bias with 95$\%$ limits of agreement from -6.1$\%$ to 5.4$\%$ compared to reference values from commercial software. This range falls within the interobserver variability reported in previous studies. The limits of agreement were wider for both CoTracker2 and EchoTracker, worse than the interobserver variability. At inference, MyoTracker used 67$\%$ less GPU memory than CoTracker2 and 84$\%$ less than EchoTracker on large sequences (100 frames). MyoTracker was 74 times faster during inference than CoTracker2 and 11 times faster than EchoTracker with our setup. Maintaining the entire sequence in the temporal context was the greatest contributor to MyoTracker's accuracy. Slight additional gains can be made by re-enabling iterative refinement, at the cost of longer processing time.

Figures (5)

• Figure 1: Deep learning approaches as alternatives to traditional STE. Point tracking and optical flow estimation are pure tracking methodologies for learning motion patterns, whereas segmentation methods detect specific structure(s) in every frame. Obtaining the myocardial outline in every frame of the video with any of the approaches allows strain calculations. The flexibility of point tracking methods allows for tracking only the desired points over variable-length sequences.
• Figure 2: MyoTracker architecture. Highlighted in red are the elements that were altered from the definition of CoTracker2 karaev_cotracker_2023 to produce MyoTracker. Where possible, the original values are highlighted in blue. The point locations are predicted directly for all frames in the sequence, without iterative refinement. All CoTracker2 components pertaining to visibility estimation, sliding window processing, transformer input encoding, and virtual tracks were removed. These components are not included in the figure.
• Figure 3: Data flow and processing steps. Processing RV-focused recordings with the AFI RV tool yields single-cycle recordings, RV myocardium keypoints, and corresponding RV FWS values. The software does not disclose the steps it takes to calculate RV FWS from the keypoints.
• Figure 4: A qualitative look at the models' predictions. The left column shows sequence frames close to the start (first end-diastole), and the other columns display the last frame (second end-diastole) predictions for the same sequence using each of the models. The cases are sorted by the average trajectory error (in pixels) in MyoTracker predictions. End trajectory error and drift are provided for each prediction. Support lines are drawn at equivalent locations between the frames to illustrate potential drift after a full cardiac cycle.
• Figure 5: Peak systolic RV FWS estimation accuracy with the trained models and our strain calculation algorithm when compared to the reference values. The differences in both tracking and strain calculation methods lead to compounding errors.