Plane-Wave Ultrasound Enhancement Using Deformable Spatiotemporal Cross-Attention for Ultrafast Ultrasound Imaging (UFUS)
Abstract
Purpose
Ultrafast ultrasound (UFUS) enables imaging of blood flow, functional dynamics, cardiac motion, and real-time treatment verification. UFUS is often acquired using high frame-rate single plane-wave imaging (SPWI), which has limited image quality due to the single-angle scan. Coherent plane-wave compounding (CPWC) improves contrast and resolution by combining multiple angles, at the cost of reduced temporal resolution. This work aims to recover CPWC-like image quality from SPWI by leveraging inter-frame redundancies with attention-based temporal aggregation.
Methods
We propose an encoder–attention–decoder network that reconstructs a target frame from a short temporal window of consecutive SPWI frames. Each frame is processed by a shared multi-scale encoder to extract hierarchical features. Temporal information is integrated using motion-guided deformable cross-attention, where neighboring-frame features are aligned to the target frame via deformable sampling and then fused through attention. Ultrasound data were acquired using a P4-2v probe (2.72 MHz center frequency, ~75% bandwidth) at up to 1500 fps in wire, bubble, and human carotid scans. The dataset contains 41 scans (750 frames each; 30,750 sequences). SPWI is used as input and CPWC compounded from 41 steering angles serves as ground truth. The network is trained with a hybrid loss (MSE + perceptual loss) using a 31/5/5 scan split for training/validation/testing.
Results
The proposed method improves SPWI reconstruction quality toward CPWC references, enhancing structural fidelity and reducing noise while maintaining single-frame imaging speed. The method achieved RMSE = 0.03, PSNR = 30.56 dB, and SSIM = 0.91, outperforming SPWI (RMSE = 0.06, PSNR = 26.38 dB, SSIM = 0.85). Inference time is 0.1 s/frame, supporting real-time UFUS workflows.
Conclusion
By integrating motion-guided deformable cross-attention, the proposed method effectively exploits inter-frame redundancy to achieve high-quality reconstruction from SPWI, offering a practical solution to improve ultrafast ultrasound image quality without sacrificing frame rate.