Real-time Ultrasound Signals Processing: Denoising and Super-resolution

Ultrasound acquisition is widespread in the biomedical field, due to its properties of low cost, portability, and non-invasiveness for the patient. The processing and analysis of US signals, such as images, 2D videos, and volumetric images, allows the physician to monitor the evolution of the patient's disease, and support diagnosis, and treatments (e.g., surgery). US images are affected by speckle noise, generated by the overlap of US waves. Furthermore, low-resolution images are acquired when a high acquisition frequency is applied to accurately characterise the behaviour of anatomical features that quickly change over time. Denoising and super-resolution of US signals are relevant to improve the visual evaluation of the physician and the performance and accuracy of processing methods, such as segmentation and classification. The main requirements for the processing and analysis of US signals are real-time execution, preservation of anatomical features, and reduction of artefacts. In this context, we present a novel framework for the real-time denoising of US 2D images based on deep learning and high-performance computing, which reduces noise while preserving anatomical features in real-time execution. We extend our framework to the denoise of arbitrary US signals, such as 2D videos and 3D images, and we apply denoising algorithms that account for spatio-temporal signal properties into an image-to-image deep learning model. As a building block of this framework, we propose a novel denoising method belonging to the class of low-rank approximations, which learns and predicts the optimal thresholds of the Singular Value Decomposition. While previous denoise work compromises the computational cost and effectiveness of the method, the proposed framework achieves the results of the best denoising algorithms in terms of noise removal, anatomical feature preservation, and geometric and texture properties conservation, in a real-time execution that respects industrial constraints. The framework reduces the artefacts (e.g., blurring) and preserves the spatio-temporal consistency among frames/slices; also, it is general to the denoising algorithm, anatomical district, and noise intensity. Then, we introduce a novel framework for the real-time reconstruction of the non-acquired scan lines through an interpolating method; a deep learning model improves the results of the interpolation to match the target image (i.e., the high-resolution image). We improve the accuracy of the prediction of the reconstructed lines through the design of the network architecture and the loss function. %The design of the deep learning architecture and the loss function allow the network to improve the accuracy of the prediction of the reconstructed lines. In the context of signal approximation, we introduce our kernel-based sampling method for the reconstruction of 2D and 3D signals defined on regular and irregular grids, with an application to US 2D and 3D images. Our method improves previous work in terms of sampling quality, approximation accuracy, and geometry reconstruction with a slightly higher computational cost. For both denoising and super-resolution, we evaluate the compliance with the real-time requirement of US applications in the medical domain and provide a quantitative evaluation of denoising and super-resolution methods on US and synthetic images. Finally, we discuss the role of denoising and super-resolution as pre-processing steps for segmentation and predictive analysis of breast pathologies.