PROJECT SUMMARY Measurement of blood oxygen (O2) saturation, the fraction of oxygen-saturated hemoglobin in blood, provides information on whole-body and organ-specific O2 delivery and consumption and is used to guide therapy and intervention. Blood sampling and analysis by invasive catheterization performed under X-ray fluoroscopic guidance is the standard method used to measure O2 saturation in multiple anatomical locations in the cardiac chambers and major blood vessels. Non-invasive measurement of O2 saturation using magnetic resonance (MR) imaging was first proposed nearly 30 years ago; however, previous techniques have relied on fitting the Luz- Meiboom model and other model variants using traditional linear and non-linear regression model approaches. Although the model captures the basic underlying biophysical principles, it does not fully characterize the complex relationship between blood O2 saturation and the MR signal. Despite being a non-invasive, non- radiating alternative to invasive catheterization, the low accuracy of MR oximetry, due to inadequacy of the model as well as estimation methods, have prevented the technique from gaining clinical acceptance. We propose to overcome this critical limitation by meeting our overall objective; to deploy a model-free approach based on machine learning (ML) to develop and implement an accurate, clinically feasible, MR oximetry technique. We hypothesize that ML algorithms provide greater flexibility in parameter estimation than traditional methods, and can be trained to learn and map the true in vivo relationship that describes the sensitivity of MR blood signal to O2 saturation. We intend to achieve our objective through the following specific aims. In Aim 1, we will develop a supervised ML algorithm for MR oximetry. Pre-training will occur with training data simulated using the L-M model and then augmented with in vivo data via transfer learning. Simultaneously, in Aim 2, we will design and implement a 3D MR oximetry method for volumetric data acquisition. A volumetric map will facilitate O2 saturation measurement throughout the vascular system, and will support the combination with 4D flow to evaluate O2 delivery and consumption. In Aim 3, we will validate the proposed ML-based 3D MR oximetry technique in a small cohort of patients referred for catheter-based O2 saturation measurement. For the first time, our proposed work will apply machine learning to accurately characterize the in vivo sensitivity of the transverse relaxation time (T2) weighted MR blood signal to O2 saturation, using a unique combination of simulated and in vivo training data. ML-based MR oximetry will provide the accuracy of measurement required for clinical use, and therefore will be able to replace or reduce the frequency and duration of an invasive, radiation-based method with a safe, non-invasive alternative. ML-based MR oximetry as an imaging tool is expected to significantly improve the diagnostic value of an MR exam, and will be especially valuable in the management of patients with congenital heart disease. Our work thus aligns with the mission of NIBIB to have a positive impact on human health with the development of novel technology.

PROJECT NARRATIVE The proposed work is relevant to public health because it will lead to a non-invasive, non-radiating biomedical technology to measure blood oxygen saturation at multiple locations within the heart and other blood vessels as an alternative to invasive catheterization. The project aligns with NIBIB’s goal to improve disease detection and management as this is the first application of machine learning to refine a novel 3D volumetric magnetic resonance technique to accurately determine oxygen saturation in the blood. This will significantly enhance the capabilities of MRI to provide comprehensive disease diagnosis and assessment.