Artificial Intelligence Driven Platform for PET/MR Imaging

该项目将开发和测试一种新的患者数据驱动的人工智能

PROJECT SUMMARY This project will develop and test a novel patient data-driven artificial intelligence (AI)-assisted image processing platform to enable dose-reduced multimodal parametric simultaneous positron emission tomography/magnetic resonance imaging (PET/MRI). Data driven multiparametric imaging is an important approach in precision medicine with the state-of-the-art whole-body hybrid PET/MRI technology playing increasing roles. However, the broader application and promising potential of the has yet to be realized. Radiation exposure from the PET radiotracers remains to be a concern when using PET is used repeatedly in follow-up examinations, especially for pediatric patients. Corrections for PET attenuation and scatter using non- CT images from MRI for PET image reconstruction and quantification needs to run additional non-diagnostic MRI sequences to derive PET attenuation correction (AC) maps, not only making the scan time longer but also containing image artifact, especially when used in whole-body imaging applications. Collecting quantitative and functional image measurements with additional MRI sequences for parametric diagnostic information can further extend the scan time, making PET/MRI scans often intolerable by pediatric, elderly and motion-prone patients. With the proposed AI frameworks, we will perform MRI-based attenuation correction (MbAC), not just limited in brain but whole body which has not been solved, using patient-specific diagnostic MRI data that are collected on the integrated PET/MRI system and to generate diagnostic-quality full-dose-equivalent PET images from low- dose data (e.g., a fraction of standard dose of the radioactive tracer). Furthermore, high-resolution and multi- orientation MR images can be “synthesized” from low-resolution and noisy images collected from fast MRI scans using the developed AI frameworks. Building on our recent development of deep learning algorithms for MbAC and synthesizing MRI, CT and PET images from various types of source images, we will: 1) develop and optimize AI frameworks to generate desired high-resolution MR images with different contrasts from low-resolution data collected by rapid MRI scans; 2) develop and refine AI-driven MbAC method using synthesized high-resolution MR images and MRI-aided synthesis of full-dose-equivalent PET images from low-dose data; and 3) determine and evaluate the performance of developed AI-driven low-dose and fast PET/MRI platform in a cohort of patients who have received standard of care PET/CT. We will use quantitative image quality metrics and expert-review to compare AI-generated full-dose-equivalent images with those of the “ground truth” full-dose PET/MRI and matching PET/CT exams. This innovative low-dose and fast PET/MRI imaging approach can be implemented in clinical settings with high efficiency, reduced cost and better patient experience, especially for those who cannot use the standard of care PET/CT or PET/MRI procedures, enabling the multimodal molecular imaging applications of PET/MRI in disease staging, therapy monitoring and follow-up of patients with oncological, cardiac and neurological diseases.