Development of advanced image reconstruction technology based on deep neural network for efficient B-mode ultrasound imaging

Abstract

Ultrasound (US) is one of the most versatile medical imaging modalities. Due to its fast frame rate and radiation-free nature, it is the first choice for applications such as fetal imaging and cardiac imaging. Reconstruction of high-quality images from a limited number of radio-frequency (RF) measurements is highly desired in portable, three-dimensional, and ultrafast ultrasound imaging systems. Unfortunately, a standard beamformer produces images with limited measurements that are unsuitable for diagnostic purposes. Towards this end, a deep learning-based beamformer is proposed to generate significantly improved images under widely varying measurement conditions and channel subsampling patterns. In particular, a deep neural network is designed to directly process full or sub-sampled radio-frequency (RF) data acquired at various subsampling rates and detector configurations so that it can generate high-quality ultrasound images using a single beamformer. Unfortunately, a separate beamformer should be trained and stored for each application, demanding significant resources such as training data and memory. Accordingly, a switchable and tunable deep beamformer is proposed that can switch between various types of output such as the DAS, MVBF, DMAS, and GCF, and also adjust speckle noise removal levels at the inference phase by using a simple switch or tunable nozzle.

The aforementioned methods rely on the paired RF dataset for supervised training. In many real-world imaging situations, access to RF data is limited and acquisition of paired images is infeasible. For example, to improve the visual quality of the US images acquired using a low-cost imaging system, one needs to scan the same part using a high-end machine, which is practically impossible. Inspired by the recent theory of unsupervised learning, the applicability of optimal transport driven CycleGAN (OT-CycleGAN) is investigated for the US artifact removal problems without matched reference data. Two types of OT-CycleGAN approaches were employed

one with the partial knowledge of image degradation physics and another with the lack of such knowledge. Various US artifact removal problems are then addressed using the two types of OT-CycleGAN. Experimental results for various unsupervised US artifact removal tasks confirmed that the proposed unsupervised learning method delivers results comparable to supervised learning in many practical applications.

In conclusion, the proposed deep neural network-based framework can be used to generate output for different applications. Furthermore, the proposed method can be trained in supervised and unsupervised fashion eliminating the need for paired training data. Therefore, it can be an important platform for efficient B-mode US imaging.