Lightweight Physics-Aware Zero-Shot Ultrasound Plane-Wave Denoising

Technical Summary: Lightweight Physics-Aware Zero-Shot Ultrasound Plane-Wave Denoising

Problem Statement
Ultrasound Coherent Plane-Wave Compounding (CPWC) improves image contrast by combining echoes from multiple steered transmissions. However, increasing the number of steering angles to enhance quality significantly reduces the imaging frame rate and can introduce motion-related blurring in dynamic scenarios. Consequently, low-angle CPWC acquisitions are often necessary for high frame rates but remain highly susceptible to incoherent noise and artifacts. While deep learning-based denoising has shown promise, supervised approaches require large-scale paired datasets of noisy and clean images, which are rarely available in clinical settings due to the impossibility of acquiring truly noise-free ground truth. Existing self-supervised methods often rely on spatial downsampling or masking strategies that may not fully exploit the specific physical properties of ultrasound acquisition, and many still require extensive training data or domain-specific fine-tuning.

Methodology
The authors propose a lightweight physics-aware zero-shot denoising framework that operates directly on the test sample without external training datasets or clean reference images. The core innovation lies in leveraging the physical properties of plane-wave acquisition to generate self-supervised training pairs:

1. Physics-Aware Angular Partitioning: Instead of spatial downsampling, the method partitions the available steering angles of a low-angle compounded image into two disjoint subsets (e.g., odd-indexed and even-indexed angles).
2. Pseudo-Pair Generation: Each subset is used to reconstruct a compounded image. While both images share the same underlying anatomical structures, they possess different angle-dependent artifacts and incoherent noise components due to the reduced number of transmissions in each subset.
3. Self-Supervised Residual Learning: A lightweight Convolutional Neural Network (CNN) is trained directly on the test image using these two reconstructed subsets as pseudo-pairs. The network learns to estimate the inconsistent components (noise and artifacts) between the subsets.
4. Loss Functions: The training utilizes a symmetric residual loss to enforce bidirectional consistency between the subsets and a gradient-based consistency loss to ensure the denoised output preserves the structural edges and texture of the original input, preventing over-smoothing.
5. Architecture: The network is designed for efficiency, consisting of only two $3 \times 3$ convolutional layers with 48 channels and a final $1 \times 1$ convolution, totaling fewer than 22,000 parameters.

Key Contributions

• Zero-Shot Framework: Introduction of a denoising framework for CPWC that eliminates the need for external training data or clean reference images.
• Physics-Aware Strategy: A novel angular partitioning strategy that exploits angle-dependent artifact variations while preserving consistent tissue structures, allowing the network to distinguish anatomy from noise.
• Lightweight Design: Development of an efficient CNN architecture with only two convolutional layers, enabling computationally inexpensive training directly on the test sample.
• Comprehensive Evaluation: Extensive validation across simulation, phantom, and in vivo datasets, demonstrating performance against classical, supervised, and other self-supervised baselines.

Experimental Results
The method was evaluated on three datasets:

1. Simulation (PICMUS): The proposed method achieved the highest Contrast-to-Noise Ratio (CNR) and generalized CNR (gCNR) across all cyst rows, outperforming supervised models (MSLAE, DPWDPM) and self-supervised baselines (ZS-N2N, N2S, S2S).
2. Phantom Data: The framework demonstrated superior performance in preserving cyst boundaries and homogeneity, achieving gCNR and CNR values closest to the high-quality 75-angle reference, surpassing BM3D and other learning-based methods.
3. In Vivo Data (Carotid Artery): On both cross-sectional (CC) and longitudinal (CL) views, the method yielded the highest contrast metrics and preserved vessel wall sharpness better than competing approaches. Notably, on the CL dataset, the proposed method achieved a gCNR of 0.88 and CNR of 3.6 dB, exceeding even the 75-angle compounded reference.

In all experiments, the method passed the Kolmogorov–Smirnov (KS) test, confirming that the statistical properties of the speckle pattern were preserved and no artificial textures were introduced. While the current implementation takes approximately 6 seconds per image (including training and inference) on an NVIDIA RTX 4090, the authors note this is not yet optimized for real-time speed.

Significance and Claims
The paper claims that this approach offers a robust, data-efficient solution for high-quality ultrasound reconstruction, particularly in scenarios with limited acquisition angles or a lack of training data. By training directly on the test image, the method avoids the domain shift issues common in supervised models that rely on synthetic or phantom data. The authors emphasize that their physics-aware pairing strategy allows the network to effectively separate anatomical information from inconsistent noise without requiring domain-specific fine-tuning. While the current architecture is not optimized for real-time deployment, the authors suggest that pretraining on diverse datasets could further reduce the required training iterations, paving the way for online clinical application. The work establishes a promising path for efficient ultrasound denoising that bridges the gap between classical signal processing and deep learning without the burden of data collection.