Respiratory Signal Estimation for Free-hand 2D Ultrasound Image via Heterogeneous Alignment and Conditional Generative Learning.

Objective: 2D ultrasound sequence captures respiration-induced morphological variation of organs and is the most widely used for estimating the signal that represents the respiratory motion. However, the spatial motion introduced by clinical free-hand ultrasound acquisition mixes with respiratory motion, reducing the estimation accuracy. This study proposes an unsupervised respiratory signal estimation method based on heterogeneous information alignment to address spatial motion interference caused by free-hand acquisition.

Methods: A heterogeneous graph of ultrasound slices is created to isolate organ respiratory motion and ultrasound probe spatial motion. The hierarchical attention aggregation is employed to learn respiratory and spatial correlation independently, mapping slices into a unified respiratory feature space. Then, the task of respiratory signal estimation is reframed as a mapping learning problem, translating the image into the unified respiratory motion feature space through conditional generative learning. The respiratory signal is the feature of the slice in this unified space.

Results: The correlation between the estimated results of the proposed method and the ground truth in various free-hand acquisition modes exceeded 93%, reaching up to 97%. The time required by the model to estimate respiratory signal for a single slice is roughly 2 ms.

Conclusion: The results demonstrate that the proposed method has higher accuracy and robustness than other methods.

Significance: The proposed unsupervised method, without the constraint of fixed image acquisition positions, is better suited to ultrasound imaging and provides the foundation for real-time 4D respiratory ultrasound imaging.