[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Ben Luijten (Eindhoven University of Technology)
Boudewine Ossenkoppele (Delft University of Technology)
Nico de Jong (Delft University of Technology)
Martin Verweij (Delft University of Technology)
Massimo Mischi (Eindhoven University of Technology)
Ruud van Sloun (Eindhoven University of Technology)

Abstract:
Ultrasound imaging is an attractive imaging modality due to its low-cost and real-time feedback, but often lacks in image quality as compared to MRI and CT imaging. Conventional ultrasound image reconstruction, such as Delay-and-Sum (DAS) beamforming, aims to find an optimal reconstruction x from measurements y, and is derived from maximum-likelihood (ML) estimation of the probability distribution p(y│x). As such, no prior information is exploited in the image formation process, which limits potential image quality. Maximum-a-posteriori (MAP) beamforming techniques aim to overcome this issue by including a prior distribution p(x). However, such methods often rely on rough approximations of the underlying signal statistics [1], or are prohibitively slow and complex for real-time imaging [2]. Deep learning based reconstruction methods have demonstrated impressive results over the past years, but often lack interpretability and require vast amounts of training data. Recently we have proposed a novel method, neural MAP, which efficiently combines deep learning in the MAP beamforming framework. In this framework we aim to overcome the challenges in MAP beamforming by learning the measured signal statistics, and the prior signal distribution, through neural networks. We show that this model-based deep learning approach can achieve high-quality imaging, improving over the state-of-the-art, without compromising the real-time abilities of ultrasound imaging. We acquired a train and test set containing distinct in-vivo images from different anatomies, through a Verasonics research platform in combination with a 128-channel 6.25-MHz linear array transducer. An 11 plane-wave (PW) imaging scheme was adopted, with transmitting angles equidistantly distributed across ± 18º. From these acquisitions we generated ground-truth images using a minimum-variance beamformer [3]. In this work we aimed to reconstruct these high-quality ground truth images from only a single 0º PW acquisition. We compared neural MAP against DAS (reference) and ABLE (state-of-the-art) [4], and evaluated the robustness of each method to input noise. To that end, we varied the input SNR from 0dB to 20dB compared to the original signal. On average we measured an increase of 9.9dB and 1.1dB in PSNR using neural MAP, compared to DAS and ABLE, respectively.