Biomarkers.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Data Analysis

Background:

• Harmonizing Positron Emission Tomography (PET) imaging data, such as Pittsburgh Compound-B (PiB) and Fluorodeoxyglucose-PET (FDG-PET), is crucial for accurate analysis.
• Traditional harmonization methods often require large amounts of paired data, which can be difficult or impossible to obtain.
• Existing unpaired methods may lack the generality needed for complex medical datasets.

Purpose of the Study:

• To develop and validate a novel Cycle-GAN based harmonization model for neuroimaging data.
• To eliminate the need for extensive paired data during model training, utilizing unpaired data instead.
• To improve the generality and performance of data harmonization in PET imaging.

Main Methods:

• A Cycle-GAN model was adapted for tabular neuroimaging data, modifying generators and discriminators with multilayer perceptrons (MLP) and skip-connections.
• The model was trained using a large dataset of unpaired PiB and FBP (presumably FBP is another PET tracer or imaging modality) measurements.
• A small fraction of paired data was used for model selection, evaluating performance via Pearson correlation of mean-cortical Standardized Uptake Value Ratio (mcSUVR).

Main Results:

• The selected Cycle-GAN model achieved a high Pearson correlation (0.85) between harmonized PiB data and actual PiB data, outperforming the baseline.
• Statistical analysis (Steiger's Z test) confirmed a significant improvement (p < 0.0001) over the baseline method.
• Inclusion of additional features like CL and demographic data (age, sex) further improved harmonization results.

Conclusions:

• A Cycle-GAN based harmonization model was successfully developed for unpaired neuroimaging data.
• The model demonstrates the feasibility of training harmonization models with predominantly unpaired data, requiring only a small subset for selection.
• Promising harmonization results were achieved for PiB and FBP measurements, indicating potential for broader application in neuroimaging analysis.