Machine learning-based amide proton transfer imaging using partially synthetic training data.

Area of Science:

• Biomedical Imaging
• Machine Learning
• Quantitative MRI

Background:

• Machine learning (ML) models for chemical exchange saturation transfer (CEST) effect quantification face challenges with limited measured data or bias from fully simulated data.
• Developing robust ML models requires addressing these data limitations in CEST imaging.

Purpose of the Study:

• Introduce a novel platform for generating partially synthetic CEST data by combining simulated and measured components.
• Evaluate the feasibility of using this partially synthetic data for training ML models to predict amide proton transfer (APT) effects.

Main Methods:

• Partially synthetic CEST signals were created by combining simulated APT effects with measured components.
• ML models were trained on partially synthetic, fully simulated, and in vivo data to predict APT effects in rat brains with 9L tumors.

Main Results:

• Partially synthetic data enabled accurate APT prediction in tissue-mimicking experiments.
• In vivo experiments demonstrated superior accuracy and robustness of ML models trained with partially synthetic data compared to other methods.

Conclusions:

• Partially synthetic CEST data effectively addresses limitations of conventional ML training approaches.
• This method offers a promising solution for enhancing the accuracy and reliability of ML-based CEST quantification.