A machine learning ultrasound beamformer based on realistic wave physics for high body mass index imaging

点击翻译按钮获取中文摘要

PROJECT SUMMARY Obesity is a significant and growing problem in the United States. Currently, 68.5% of the U.S. population is overweight, with approximately 37.7% of the overweight population being obese. The significant health problems associated with overweightedness and obesity, the “body habitus” of this population combined with the significant challenges in medical imaging of these individuals reduces the effectiveness of healthcare for this population. In ultrasound imaging, the quality of abdominal ultrasound exams are significantly affected by obesity. Fundamentally, an ultrasound image relies on acoustic propagation to a target, reflection, and then propagation back to the surface. The process of beamforming, which converts the surface measurement to an image, is sensitive to the low amplitude reflections from different tissue layers and tissue properties. Typically, the additional fat and connective tissue layers in obese patients can significantly degrade ultrasound image quality by introducing multi-path reverberation and phase aberration that obscure or distort these low amplitude reflections. However, due to the computational complexity of describing ultrasound propagation and reflection in heterogeneous media, beamformers currently rely on simplified models that do not describe the propagation physics directly. We propose a generational leap in how we approach ultrasound beamforming by using physically and anatomically realistic wave propagation models and measurements that can effectively harness the power of data-driven and rapidly evolving machine learning beamformers. A custom highly realistic simulation tool that we have developed will use acoustical maps of the fine structures in the human body based on photographic cryosections. This physics-based approach will allow us to develop high quality training data and to understand the physical mechanisms for image quality improvement. These simulations will be calibrated to ex vivo and in vivo human data to subsequently generate a large data set that can be used to train a machine- learning-based real-time beamformer. We will focus on two sources of image degradation which we have identified to be particularly deleterious: multipath reverberation and aberration of the focusing profile. The proposed neural network beamformer filters incoherent noise, such as multi-path reverberation, and corrects aberration in the radiofrequency channel signals. After training the beamformer and implementing it in real-time, a pilot human study in liver ultrasound imaging will be conducted to determine the improvement in image quality in high-body-mass index individuals, where diagnostic imaging is problematic due to image degradation. This technique is highly translatable to other clinical scenarios, varying from cardiac to transcranial to obstetric imaging, by changing the anatomical model. Furthermore, the physical concepts that will be extracted from the learned representation, can be used to improve the design process for ultrasound equipment, including transmit sequences, and transducers.