Deep learning in ultrasound imaging (1907.02994v2)

Abstract: We consider deep learning strategies in ultrasound systems, from the front-end to advanced applications. Our goal is to provide the reader with a broad understanding of the possible impact of deep learning methodologies on many aspects of ultrasound imaging. In particular, we discuss methods that lie at the interface of signal acquisition and machine learning, exploiting both data structure (e.g. sparsity in some domain) and data dimensionality (big data) already at the raw radio-frequency channel stage. As some examples, we outline efficient and effective deep learning solutions for adaptive beamforming and adaptive spectral Doppler through artificial agents, learn compressive encodings for color Doppler, and provide a framework for structured signal recovery by learning fast approximations of iterative minimization problems, with applications to clutter suppression and super-resolution ultrasound. These emerging technologies may have considerable impact on ultrasound imaging, showing promise across key components in the receive processing chain.

Deep Learning in Ultrasound Imaging

This paper investigates the integration of deep learning (DL) techniques within the entire ultrasound (US) imaging chain, aiming to enhance reconstruction quality and expand diagnostic capabilities. The authors present DL applications from the front-end processing stages to advanced imaging scenarios, highlighting the necessity of embedding domain-specific knowledge to achieve efficient and effective solutions.

Ultrasound imaging remains an indispensable diagnostic tool due to its cost-effectiveness and real-time capabilities. However, it faces limitations such as penetration depth and signal quality. Recent advances have seen US systems becoming more compact and capable of 3D imaging, handling higher data rates. AI methods, especially those from DL paradigms, are progressively becoming essential in pushing the received signal processing boundaries beyond traditional methodologies.

Core Contributions

The paper's pivotal contributions involve deploying DL techniques across various aspects of US imaging and demonstrating their efficacy:

1. Beamforming and Adaptive Processing: DL models enhance traditional delay-and-sum beamforming by learning adaptive apodization weights, improving resolution and contrast. By embedding structural priors, the networks exhibit improved performance over static, hand-tuned approaches and manage patient-specific variability more effectively.
2. Spectral Doppler Advancements: By using neural networks as adaptive spectral estimators, the paper reduces spectral leakage and improves resolution compared to traditional periodogram-based methods. This approach acknowledges the power of DL to calculate position-specific filter bank parameters, optimizing for temporal and spectral resolution trade-offs.
3. Compressive Encodings: Targeting communication bandwidth constraints, DL methods are employed to encode channel data at the probe for downstream decoding, maximizing transmission efficiency without compromising velocity estimations in Doppler applications.
4. Clutter Suppression through Deep Unfolding: An algorithmic unfolding of RPCA offers robust clutter suppression, moving beyond classical SVD thresholding approaches. This method derives an improved clutter-resolved representation in microbubble imaging, enhancing contrast in US imaging.
5. Super-Resolution Ultrasound Imaging: Super-resolution methods, which rely on microbubble localization, are accelerated by DL, resulting in significantly reduced acquisition times, improved localization accuracy, and enhanced image resolution. Deep networks are leveraged as tools for robust sparse recovery to enhance ULM capabilities.

Implications and Future Outlook

The implications of these advancements are substantial for clinical applications, positioning DL as a key enabler in extracting detailed, high-quality diagnostic information from US scans. The potential for real-time adaptive processing can address individual patient variability, potentially reducing the skill threshold required for accurate US interpretation. Moreover, the proposed DL methods alleviate data transmission bottlenecks in portable devices, broadening accessibility and applicability.

Looking forward, integrating unsupervised and continual learning paradigms into these systems could allow US systems to adapt continuously to evolving clinical data. This continuous learning ability places future intelligent US probes at a unique vantage point, seamlessly merging data-driven insights with classical signal processing to innovate beyond current capabilities.

In conclusion, this paper presents substantial progress in integrating DL within US imaging, demonstrating clearly the intertwined evolution of these fields. The demonstrated methods lay a groundwork for further development of smart US systems, where DL enriches signal processing, enhances resolution, and provides actionable insights in real-time.