Advancing Machine Learning for CEST MRI Analysis through Partially Synthetic Data

Vanderbilt University Medical Center, Nashville TN

Investigators

Abstract

PROJECT SUMMARY Glioma represents a significant health concern, characterized by alterations in multiple macromolecules and metabolites within the tumor microenvironment. Non-invasive imaging of these molecular profiles could provide invaluable diagnostic insights. Chemical exchange saturation transfer (CEST) is an emerging molecular imaging technique that offers enhanced sensitivity and provides complementary information to magnetic resonance spectroscopy (MRS), with the ability to reflect various molecular compositions within tissues. Amide proton transfer (APT), a primary variant of CEST that reflects mobile proteins, has demonstrated substantial potential for glioma diagnosis. However, despite FDA approval for APT imaging in glioma in 2018, it has not yet been integrated into clinical practice. A clinical trial aimed at differentiating gliomas using APT imaging was halted due to poor image quality, likely resulting from the inaccuracy of existing CEST quantification methods and the inherently low signal-to-noise ratio (SNR) of CEST imaging. Machine learning (ML) has the potential to identify complex features that traditional methods often miss. While ML has been applied to quantify CEST, it faces challenges such as insufficient training data and poor-quality ground truth data when trained using measured in vivo data. Fully synthetic data can mitigate these issues, but practical application remains challenging due to unidentified exchangeable pools and their parameter ranges in tissues. Our project aims to develop a novel platform to generate partially synthetic CEST data for training ML models, thereby addressing these challenges. Preliminary data suggest success in accurately quantifying APT in animal models in preclinical MRI using an ideal steady-state continuous-wave saturation. In Aim 1, we will translate this approach to human imaging at 3T by extending it to the more complex non-steady-state pulsed-CEST imaging typically used in clinical MRI. We will also expand it beyond APT to include amine (glutamate), guanidine (creatine), and nuclear Overhauser enhancement (NOE) effects (large proteins, phospholipids), enabling more generalized CEST imaging of molecular profiles. In Aim 2, we will develop a novel autoencoder-based denoising technique, coupled with an innovative dual-power data preparation strategy, to improve SNR. This strategy leverages the high SNR of higher saturation power and the enhanced peak resolvability of lower power, ensuring effective noise suppression while preserving molecular information. This will make the technique more suitable for low-power applications, which are preferred for imaging multiple CEST effects at 3T. In Aim 3, we will integrate these two ML-based techniques and evaluate their performance in glioma patients. By the completion of this project, we aim to overcome current challenges in CEST imaging, facilitating its widespread clinical adoption and significantly enhancing healthcare outcomes.