Echo Flow-Induced Temporal Correlation Learning for Ultrasound Video Object Segmentation.

Objective: The segmentation of ultrasound video objects aims to delineate specific anatomical structures or areas of injury in sequential ultrasound imaging data. Current methods exhibit promising results, but struggle with key aspects of ultrasound video analysis. They insufficiently capture inter-frame object motion, resulting in unsatisfactory segmentation for dynamic or low-contrast scenarios. With the release of SAM2, video object segmentation has advanced significantly. However, its performance in ultrasound videos remains suboptimal due to its design bias toward natural videos and lack of consideration for ultrasound-specific characteristics. We propose a novel EchoSAM2 method to achieve more accurate object segmentation in ultrasound videos.

Methods: We propose Echo Flow, which captures motion trends between frames to enhance the modeling of temporal relationships. It also helps suppress interference from non-object regions by leveraging object motion patterns. Furthermore, we propose an Echo Modulation Block (EMB) to seamlessly incorporate Echo Flow into the SAM2 framework, improving the quality of feature representation. To further optimize SAM2's performance during fine-tuning, we present a Gaussian Adapter specifically designed for ultrasound image characteristics.

Results: Extensive experiments on three ultrasound video datasets confirm the effectiveness of our method, achieving state-of-the-art results. On the EUDP dataset, our model achieves a Dice of 85.49%, outperforming the second-best method by 3.19%. Models trained on HMC-QU and CAMUS achieve the best generalization when tested on each other's unseen test sets.

Conclusion: The introduction of Echo Flow, along with other supporting modules, enhances both segmentation accuracy and the model's generalizability.

Significance: Accurate segmentation of ultrasound video objects enhances diagnostic accuracy and consistency, thereby increasing overall clinical value.