Abstract: In portable, 3-D, or ultra-fast ultrasound (US) imaging systems, there is an increasing demand to reconstruct high-quality images from a limited amount of data. However, existing solutions require either hardware changes or computationally expensive algorithms. To overcome these limitations, we propose a novel deep learning approach that interpolates missing radio-frequency (RF) data by utilizing the sparsity of the RF data in the Fourier domain. Extensive experimental results from sub-sampled RF data from a real US system confirm that the proposed method can effectively reduce the data rate without sacrificing image quality.
Introduction
The advancement of medical imaging toward ultra-fast and portable 3D ultrasound systems is hindered by the massive volume of radio-frequency (RF) data generated. Capturing high-resolution images at high speeds traditionally requires significant data bandwidth. To mitigate this, researchers have explored data reduction techniques, but these often necessitate specialized hardware or complex "compressed sensing" algorithms. Mathematical approaches, such as low-rank Hankel matrix completion, are theoretically sound but computationally expensive, making them impractical for real-time clinical diagnostics. This study introduces a deep learning framework designed to achieve high-speed data recovery with the efficiency required for real-time applications.
Proposed Methodology
The proposed method utilizes a novel Convolutional Neural Network (CNN) architecture designed to interpolate missing RF data. Unlike conventional medical AI that operates on post-processed images, this network processes raw signals in the Rx-SC (Receiver-Scanline) domain.
The architecture is grounded in the Deep Convolutional Framelets framework. Key technical features include:
- RF-Domain Processing: By working on raw signals before image formation, the network preserves more information and avoids the propagation of artifacts.
- Network Architecture: A deep network consisting of 28 convolutional layers was implemented, utilizing batch normalization, ReLU activation, and residual "bypass connections" to learn the residuals between sub-sampled data and ground truth.
- Standard Backend Integration: Since the CNN reconstructs raw RF data, the output remains compatible with standard Delay-and-Sum (DAS) beamformers, allowing for seamless integration into existing ultrasound software pipelines.
Theoretical Insights
A core insight of this research is that a CNN can be interpreted as a learned, accelerated version of a Hankel matrix decomposition. The researchers leverage the inherent redundancy of ultrasound signals, which are highly "sparse" in the Fourier domain.
By performing interpolation directly in the RF domain, the system removes blurring and line artifacts before the image is formed. Furthermore, because the model learns the physics of signal redundancy rather than specific image features, it exhibits remarkable universality. A model trained on a linear array probe (typically for shallow tissue) can successfully reconstruct data from a convex array probe (used for deep organ imaging), a feat that often eludes image-domain models.
Experimental Results
The system was validated using real-world data from an E-CUBE 12R US system. Training and validation were conducted using 18,000 Rx-SC planes from linear array transducers, while universality was tested using liver scans from a convex array transducer.
- 4x Data Acceleration: The model successfully reconstructed high-quality images using only 25% of the original RF data.
- Quantitative Improvement: The CNN-based reconstruction achieved a Peak Signal-to-Noise Ratio (PSNR) of approximately 30.92 dB, representing a 10 dB improvement over standard sub-sampled reconstruction.
- Visual Fidelity: Qualitative analysis showed that the CNN effectively eliminated the blurring and artifacts typical of data skipping, producing images nearly indistinguishable from full-data acquisitions.
Discussion and Future Work
While the proposed deep learning approach demonstrates significant potential for accelerating ultrasound imaging, several areas remain for future exploration. The current study focused on random sub-sampling; future research could investigate whether optimized sampling patterns might further enhance acceleration. Additionally, while the 28-layer network is efficient, detailed hardware implementation on FPGAs or mobile GPUs is necessary for bedside deployment. Finally, formal clinical validation by radiologists is required to ensure that the interpolated data does not introduce subtle artifacts that could impact diagnostic accuracy.
Summary updated