Abstract
Accurate motion estimation for tracking deformable tissues in echocardiography is essential for precise cardiac function measurements. While traditional methods like block matching or optical flow struggle with intricate cardiac motion, modern point tracking approaches remain largely underexplored in this domain. This work investigates the potential of state-of-the-art (SOTA) point tracking methods for ultrasound, with a focus on echocardiography. Although these novel approaches demonstrate strong performance in general videos, their effectiveness and generalizability in echocardiography remain limited. By analyzing cardiac motion throughout the heart cycle in real B-mode ultrasound videos, we identify that a directional motion bias across different views is affecting the existing training strategies. To mitigate this, we refine the training procedure and incorporate a set of tailored augmentations to reduce the bias and enhance tracking robustness and generalization through impartial cardiac motion. We also propose a lightweight network leveraging multi-scale cost volumes from spatial context alone to challenge the advanced spatiotemporal point tracking models. Experiments demonstrate that fine-tuning with our strategies significantly improves models' performances over their baselines, even for out-of-distribution (OOD) cases. For instance, EchoTracker boosts overall position accuracy by 60.7% and reduces median trajectory error by 61.5% across heart cycle phases. Interestingly, several point tracking models fail to outperform our proposed simple model in terms of tracking accuracy and generalization, reflecting their limitations when applied to echocardiography. Nevertheless, clinical evaluation reveals that these methods improve GLS measurements, aligning more closely with expert-validated, semi-automated tools and thus demonstrating better reproducibility in real-world applications.
Abstract (translated)
在超声心动图中，对可变形组织进行准确的运动估计对于精确测量心脏功能至关重要。尽管传统的块匹配或光学流方法难以处理复杂的心脏运动，现代点跟踪技术在此领域的应用仍相对较少。本研究探讨了当前最先进的（SOTA）点跟踪方法在超声波中的潜力，并着重于其在超声心动图上的适用性。虽然这些新技术在一般视频中表现出色，但在超声心动图领域内的有效性和泛化能力仍然有限。 通过对整个心脏周期的真实B模式超声视频的心脏运动进行分析，我们发现现有的训练策略在不同视图之间存在方向性运动偏差，影响了模型性能。为了缓解这一问题，我们优化了训练流程，并引入了一套定制的数据增强技术以减少偏倚并提高跟踪的鲁棒性和泛化能力。此外，我们提出了一种轻量级网络，利用空间上下文中的多尺度成本体来挑战高级时空点跟踪模型。 实验结果表明，采用我们的策略进行微调可以显著提升模型在基准测试上的表现，即使对于分布外（OOD）案例也是如此。例如，EchoTracker 在心脏周期各个阶段的整体位置精度提高了60.7%，中位轨迹误差减少了61.5%。有趣的是，一些点跟踪模型未能在跟踪准确性和泛化性方面超越我们提出的简单模型，这表明它们应用于超声心动图时存在局限性。 然而，临床评估显示这些方法改善了GLS（环形长轴应变）的测量值，并且更接近专家验证的半自动工具的结果，从而在实际应用中表现出更好的可重复性。
URL
https://arxiv.org/abs/2507.10127
