# Data-Driven Image-Based Protocol for Brain PET Image Harmonization

## TL;DR

A new method harmonizes brain PET images across different scanners without using phantom scans, but its accuracy drops at higher resolutions.

## Contribution

A data-driven protocol for harmonizing brain FDG-PET images without phantom scans is introduced.

## Key findings

- The harmonization protocol achieves robust accuracy at moderate resolutions (8 and 10 mm FWHM).
- At higher resolutions (6 mm FWHM), harmonization accuracy decreases significantly.

## Abstract

What are the main findings?
A novel image-based data-driven harmonization protocol successfully estimates scanner-specific filters for brain FDG-PET, without requiring phantom scans.Harmonization accuracy is robust for moderate target resolutions (8 and 10 mm) but is notably reduced at higher resolutions (6 mm).

A novel image-based data-driven harmonization protocol successfully estimates scanner-specific filters for brain FDG-PET, without requiring phantom scans.

Harmonization accuracy is robust for moderate target resolutions (8 and 10 mm) but is notably reduced at higher resolutions (6 mm).

What is the implication of the main finding?
This method facilitates harmonization of retrospective multicenter FDG-PET brain studies, enhancing comparability even when phantom calibration data are unavailable.Limitations identified at higher resolutions highlight the need for further methodological improvements, aligning harmonization strategies with recent advancements in high-resolution PET imaging.

This method facilitates harmonization of retrospective multicenter FDG-PET brain studies, enhancing comparability even when phantom calibration data are unavailable.

Limitations identified at higher resolutions highlight the need for further methodological improvements, aligning harmonization strategies with recent advancements in high-resolution PET imaging.

Quantitative FDG-PET brain imaging across multiple centers is challenged by inter-scanner variability, impacting the comparability of neuroimaging data. This study proposes a data-driven image-based harmonization protocol to address these discrepancies without relying on traditional phantom scans. The protocol uses spatially normalized FDG-PET brain images to estimate scanner-specific Gaussian smoothing filters, optimizing parameters via the structural similarity index (SSIM). Validation was performed using images from cognitively normal individuals and Alzheimer’s disease patients from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Results demonstrated robust harmonization at moderate target resolutions (8 and 10 mm FWHM), with filter estimates consistently within 1.2 mm of phantom-derived ground truths. However, at higher resolutions (6 mm FWHM), discrepancies reached up to 3 mm, reflecting reduced accuracy. These deviations were particularly evident for high-resolution scanners like HRRT, likely due to elevated noise levels and smaller sample sizes. The presented harmonization method effectively reduces inter-scanner variability in retrospective FDG-PET studies, especially valuable when phantom scans are unavailable. Nonetheless, the current limitations at finer resolutions underline the necessity for methodological refinements to meet the demands of evolving high-resolution PET imaging technologies.