Comparative Analysis of Generalization and Harmonization Methods for 3D Brain fMRI Images: A Case Study on OpenBHB Dataset

Due to limitations on the amount of available data from a single source, combining data from different sources can significantly improve the statistical analysis of fMRI images. However, because the training and target sources are usually different, applying a deep learning model trained on the source domain leads to inconsistent results when applied to the target domain, especially on 3D MRI images due to variations in bias across scanners. While Harmonization methods and Domain Adaptation (DA) are popular approaches for handling multi-site MRI data, Domain Generalization (DG) is less studied and may offer some unique benefits.In this study, we explore the impact of DG, and compare data harmonization on brain age prediction using 3D fMRI images from the OpenBHB dataset. The dataset consists of 3D T1 brain MRI scans aggregated from 10 publicly available datasets, comprising N = 3985 individuals and acquired on more than 70 different scanners. We focus on the Big Healthy Brains (BHB) dataset, utilizing Voxel - Based Morphometry (VBM) images and performing a split into training, out-of-domain validation, and out-of-domain test sets based on site metadata. Our experiments involve training models based on the architecture introduced in [8], extending it as per [9], and evaluating various generalization techniques on both harmonized and non-harmonized data and their differences. Our findings highlight that harmonization increases the Mean Absolute Error (MAE) for all methods. DA and DG methods reliably improve performance on out-of-domain tests, with Domain Adversarial Neural Network (DANN) exhibiting the best performance. This paper reports the complex interplay between data harmonization and model generalization, providing insights into the selection and application of generalization techniques in neuroimaging.