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

PURPOSE: Center-specific effects in PET brain scans arise due to differences in technical and procedural aspects. This restricts the merging of data between centers and introduces source-specific bias.

METHODS: We demonstrate the use of the recently proposed machine learning method Iterated Relevance Matrix Analysis (IRMA) for harmonization of center-specific effects in brain 18 F-Fluorodeoxyglucose ( 18 F-FDG) PET scans. The center difference is learned by applying IRMA on PCA-based feature vectors of healthy controls (HC), resulting in a subspace V , representing information not comparable between centers, and the remaining subspace U , where no center differences are present. In this proof-of-concept study, we demonstrate the properties of the method using data from four centers. After center-harmonization, a Generalized Matrix Learning Vector Quantization (GMLVQ) model was trained to discriminate between Parkinson's disease, Alzheimer's disease and Dementia with Lewy Bodies.

RESULTS: At the initial IRMA iteration, the system was able to determine the center origin of the four HC cohorts almost perfectly. The method required six iterations, corresponding to a six-dimensional subspace V , to determine the entire center difference. An uncorrected disease classification model was highly biased to center-specific effects, creating a falsely inflated performance when applying internal (cross-) validation. The cross-validation performance of the center-harmonized model remained high, while it generalized significantly better to unseen test cohorts. Furthermore, the framework is highly transparent, providing analytic reconstructions of the correction and visualizations of the data in voxel space.

CONCLUSION: IRMA can be used to learn and disregard center-specific information in features extracted from brain 18 F-FDG PET scans, while retaining disease-specific information.