IRMA: Machine learning-based harmonization of 18 F-FDG PET brain scans in multi-center studies

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Data Analysis

Background:

• Center-specific variations in PET scans hinder multi-center data analysis.
• Technical and procedural differences introduce bias, limiting data integration.

Purpose of the Study:

• To demonstrate Iterated Relevance Matrix Analysis (IRMA) for harmonizing center-specific effects in brain 18F-FDG PET scans.
• To improve the accuracy of disease classification models using harmonized PET data.

Main Methods:

• Applied IRMA to PCA-based feature vectors from healthy controls to identify and isolate center-specific information.
• Trained a Generalized Matrix Learning Vector Quantization (GMLVQ) model on harmonized data to classify Parkinson's disease, Alzheimer's disease, and Dementia with Lewy Bodies.
• Utilized a six-dimensional subspace to represent and remove entire center differences.

Main Results:

• IRMA effectively identified center origins of PET data within six iterations.
• Harmonized models showed high cross-validation performance and improved generalization to unseen data.
• The framework provided transparent analytic reconstructions and visualizations of the harmonization process.

Conclusions:

• IRMA successfully learns and removes center-specific information from brain 18F-FDG PET scans.
• Disease-specific information is retained, enhancing the reliability of multi-center PET data analysis.
• This method facilitates more robust and generalizable findings in neurodegenerative disease research.