Objectives

• Proposed Iterated Relevance Matrix Analysis (IRMA), a machine learning method, for harmonizing center-specific effects in multi-center 18F-Fluorodeoxyglucose (18F-FDG) PET brain scans.
• Demonstrated that IRMA can effectively remove center-specific information while retaining disease-specific information.
• Showed that IRMA harmonization improves cross-center classification performance of neurodegenerative diseases (Parkinson's disease (PD), Alzheimer's disease (AD), and Dementia with Lewy Bodies (DLB)).
• Provided analytical reconstructions of the correction and visualizations of the data in voxel space, offering transparency to the harmonization process.

Methodology

• Applied Iterated Relevance Matrix Analysis (IRMA), a machine learning method, to remove center-specific effects. IRMA iteratively trains a Generalized Matrix Learning Vector Quantization (GMLVQ) model to classify the center origin of healthy controls (HCs).
• Used Principal Component Analysis (PCA) for feature extraction from 18F-FDG PET scans.
• Accumulated center-discriminative vectors and retrained the GMLVQ model while projecting out previously found discriminative directions.
• Continued iteration until no further meaningful separation could be found between HC groups from different centers.
• Trained a GMLVQ model to classify three disease classes (PD, AD, and DLB) after center harmonization.

Results

• Center origin of four HC cohorts could be determined almost perfectly before harmonization (Balanced Accuracy (BAC) = 0.72, Area Under the Curve (AUC) = 0.89).
• Six IRMA iterations were required to remove all center-specific information (BAC reached random performance).
• Cross-validation performance of the center-harmonized disease classification model remained high (BAC = 0.69, AUC = 0.84) compared to the uncorrected model (BAC = 0.78, AUC = 0.94).
• Cross-center classification performance on unseen test cohorts significantly improved after IRMA harmonization (BAC increased from 0.41 to 0.59, AUC increased from 0.55 to 0.79).
• IRMA harmonization outperformed center-wise z-scoring and showed a slight advantage over ComBat harmonization.

Discussions

• The study demonstrates a novel application of IRMA for harmonizing multi-center 18F-FDG PET data, showing promising results. However, the reliance on HC cohorts for harmonization is a limitation, as acquiring HC data can be challenging and costly.
• The study used a relatively small number of centers (four). Further validation with a larger number of centers and more diverse datasets is needed.
• While the authors compared IRMA to ComBat and center-wise z-scoring, a more comprehensive comparison with other harmonization methods, including image-level harmonization techniques, would strengthen the findings.
• The study acknowledges potential clinical differences between cohorts (e.g., disease duration). A more detailed analysis of the impact of these differences on the harmonization results would be beneficial.
• The authors mention future work to estimate the minimum number of HCs required. Providing a more concrete guideline or power analysis for sample size determination would be valuable.

---

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