Mitigating MRI Domain Shift in Sex Classification: A Deep Learning Approach with ComBat Harmonization

TL;DR

This study addresses the challenge of domain shift in MRI-based sex classification by applying ComBat harmonization, significantly improving cross-dataset model performance and emphasizing the importance of domain adaptation in neuroimaging AI.

Contribution

The paper demonstrates that ComBat harmonization effectively reduces domain shift in MRI data, enhancing the generalizability of deep learning models for sex classification across different datasets.

Findings

ComBat harmonization improves cross-domain accuracy from 0.50 to 0.61.

Deep learning models perform well within datasets but poorly across datasets without harmonization.

Feature visualization confirms reduced domain discrepancy after harmonization.

Abstract

Deep learning models for medical image analysis often suffer from performance degradation when applied to data from different scanners or protocols, a phenomenon known as domain shift. This study investigates this challenge in the context of sex classification from 3D T1-weighted brain magnetic resonance imaging (MRI) scans using the IXI and OASIS3 datasets. While models achieved high within-domain accuracy (around 0.95) when trained and tested on a single dataset (IXI or OASIS3), we demonstrate a significant performance drop to chance level (about 0.50) when models trained on one dataset are tested on the other, highlighting the presence of a strong domain shift. To address this, we employed the ComBat harmonization technique to align the feature distributions of the two datasets. We evaluated three state-of-the-art 3D deep learning architectures (3D ResNet18, 3D DenseNet, and 3D EfficientNet) across multiple training strategies. Our results show that ComBat harmonization effectively reduces the domain shift, leading to a substantial improvement in cross-domain classification performance. For instance, the cross-domain balanced accuracy of our best model (ResNet18 3D with Attention) improved from approximately 0.50 (chance level) to 0.61 after harmonization. t-SNE visualization of extracted features provides clear qualitative evidence of the reduced domain discrepancy post-harmonization. This work underscores the critical importance of domain adaptation techniques for building robust and generalizable neuroimaging AI models.