Enhancing amide proton transfer imaging in ischemic stroke using a machine learning approach with partially synthetic...

Area of Science:

• Biomedical Imaging
• Medical Physics
• Machine Learning Applications

Background:

• Amide proton transfer (APT) imaging is sensitive to tissue pH and crucial for diagnosing ischemic stroke.
• Conventional APT quantification methods are often inaccurate or time-consuming.
• Machine learning (ML) offers a potential solution for improving APT quantification accuracy and speed.

Purpose of the Study:

• To develop and evaluate an ML model for accurate and rapid APT imaging in an animal stroke model.
• To investigate the efficacy of partially synthetic data for training ML models in this context.
• To compare the ML approach with conventional APT quantification techniques.

Main Methods:

• An ML model was trained on a novel type of partially synthetic data, integrating measured and simulated chemical exchange saturation transfer (CEST) components.
• Recursive feature elimination was used for optimization, reducing selected frequency offsets from 69 to 13.
• The ML model's performance was compared against conventional methods like asymmetric analysis and Lorentzian fit.

Main Results:

• The ML model trained on partially synthetic data accurately predicted APT effects, clearly delineating stroke lesions from normal tissue.
• This approach showed significantly enhanced contrast compared to conventional quantification methods.
• ML models trained on fully synthetic or in vivo data demonstrated poorer predictive performance.
• Optimization significantly reduced scan time by selecting fewer frequency offsets.

Conclusions:

• Partially synthetic data provides a practical and effective approach for training ML models for APT imaging.
• The developed ML method offers a more accurate and faster alternative to conventional techniques for stroke diagnosis.
• This ML-based strategy holds significant promise for improving clinical applications of APT imaging in ischemic stroke detection.