CDS&E: Collaborative Research: Scalable Deep Learning-Based Quantitative Ultrasound Tomography

点击翻译按钮获取中文摘要

Empowered by advanced imaging methods, ultrasound computed tomography (USCT) provides highly specific tissue differentiation and diagnosis through rapid, low-cost scanning without using sedatives and X-rays. However, these physics-based imaging methods currently require a long processing time and at times encounter reconstruction errors, hindering them from being widely used in time-sensitive USCT applications (e.g., brain and breast imaging). To address these gaps, this project will create two artificial intelligence-based approaches to significantly improve the reconstruction speed and the quality of the state-of-the-art USCT technologies. A faster and more powerful USCT system will potentially lead to societal benefits such as improved medical outcomes, reduced risks, higher patient satisfaction, and reduced healthcare costs. The collaborative team of investigators will actively recruit and mentor undergraduate and graduate students with diverse backgrounds to participate in this research, reach out to K-12 educators and students to create student interest in data science, computing, and STEM fields, incorporate the findings into class modules, and disseminate the technology and findings to the public.This goal of this project is to create two open-source, high-performance computing (HPC)-enabled, and deep learning (DL)-based frameworks to significantly improve the reconstruction speed and the quality of full waveform inversion (FWI)-based USCT. FWI techniques have recently enabled USCT in the reconstruction of detailed quantitative material/tissue parameters, but they have a slow reconstruction speed and a high-computational cost. To address these challenges, one of the approaches will innovatively incorporate the adjoint-tomography theory (ATT) into a generative adversarial network (GAN) to reliably accelerate FWI-based USCT by providing strong priors for GAN. Rapid patient screening can be achieved using this method. The second DL approach leverages physics-guided, cycle-consistence in both training and its application to provide extraordinary, detailed reconstruction. The second method reduces the reliance on ground truth models in training, alleviates the dependence of initial models, and utilizes the purposely-built computing hardware for DL acceleration and hence can a) lower the false-positive rates while reducing unnecessary extra tests/biopsies and b) decrease the false negatives to enable early diagnosis/treatment. The theoretical foundations will be derived for both approaches. Computational details such as framework designing/tuning and the studies of scientific problems (e.g., influences of initial models and forward modeling errors) will be generated and disseminated. A cyberinfrastructure that provides HPC/GPU-enabled imaging data generation and training frameworks and two clinically relevant databases suitable for DL-based USCT studies will be created and made available to the public. The framework building philosophy and approaches developed in this project will demonstrate an efficient, systematic approach to applying deep learning-based techniques in a scalable and parallel manner, thereby accelerating broader DL-related topics in ultrasound imaging, photoacoustic tomography, X-ray computed tomography, radar technologies, geophysics, and magnetic resonance imaging.This project is jointly funded by the Engineering of Biomedical Systems (EBMS) Program and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.