Developing a Novel Machine Learning Method with Partially Synthetic Data for Rapid and Precise pH Mapping in Ischemic Stroke

Vanderbilt University Medical Center, Nashville TN

Investigators

Abstract

PROJECT SUMMARY Neuroimaging, particularly for assessing salvageable tissue, is vital to modern stroke diagnosis and treatment decisions. It has been previously proved that pH/diffusion mismatch provides more accurate delineation of penumbra than the conventional perfusion/diffusion mismatch. This project aims to advance an MRI pH imaging technique, amide proton transfer (APT), for detecting penumbra. Despite APT's sensitivity to pH, quantifying its effect in tissues is challenging due to several confounding factors. Current techniques either fail to eliminate these contaminations or require lengthy scan times. As a result, although APT has showed potential in diagnosing stroke patients for a decade, it has not yet been adopted in clinical practice. Recently, machine learning (ML) has shown promise in improving the quantification of APT effects by identifying complex features that traditional methods often miss. However, ML often face issues such as insufficient training data and poor-quality ground truth when trained using in vivo data. This problem is exacerbated in acute stroke where obtaining large and high-quality training datasets is difficult. While fully synthetic data can address these issues, practical application remains challenging due to unidentified exchangeable pools and their parameter ranges. Our project proposes developing a new platform that generates partially synthetic chemical exchange saturation transfer (CEST) data for training ML models, addressing these challenges. Unlike a simple blend of measured and simulated CEST signals, the partially synthetic data integrates underlying components including CEST, direct water saturation, and magnetization transfer effects derived from either measurements or simulations to reconstruct CEST Z-spectrum. This approach balances simulation flexibility and fidelity. Recently, we published this novel concept of partially synthetic CEST data for ML training under the ideal condition of steady-state continuous-wave saturation achievable in preclinical MRI. Most recently, we published its application in an animal stroke model which outperformed traditional APT quantification methods and other ML methods trained on either in vivo or fully synthetic data in accurately and robustly detecting stroke. Additional experiments on healthy humans using steady-state pulsed saturation at 3T MRI demonstrate the method's transferability to human imaging. This proposal aims to further develop the method by extending it to the more complex non-steady-state pulsed saturation commonly used in clinical MRI, implementing it with interleaved acquisition and optimizing frequency offsets to shorten scan times, thus facilitating its translation to stroke patients. We will also validate its specificity and sensitivity through simulations and control phantoms in Aim 1, assess its ability to detect penumbra in an animal stroke model in Aim 2, and implement it at 3T MRI for a pilot study in human patients. Upon completion, this project will pave the way for the widespread adoption of APT imaging in identifying penumbra outside traditional treatment windows, significantly impacting healthcare.