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

# 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. ![](https://media.springernature.com/lw1200/springer-static/image/art%3A10.1007%2Fs00259-025-07114-4/MediaObjects/259_2025_7114_Fig2_HTML.png)
- 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. ![](https://media.springernature.com/lw1200/springer-static/image/art%3A10.1007%2Fs00259-025-07114-4/MediaObjects/259_2025_7114_Fig5_HTML.png)

# 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. ![](https://media.springernature.com/lw1200/springer-static/image/art%3A10.1007%2Fs00259-025-07114-4/MediaObjects/259_2025_7114_Fig1_HTML.png)
- 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). ![](https://media.springernature.com/lw1200/springer-static/image/art%3A10.1007%2Fs00259-025-07114-4/MediaObjects/259_2025_7114_Fig4_HTML.png)
- 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](https://link.springer.com/article/10.1007/s00259-025-07114-4)