Abstract
There is significant interest in pooling magnetic resonance image (MRI) data from multiple datasets to enable mega-analysis. Harmonization is typically performed to reduce heterogeneity when pooling MRI data across datasets. Most MRI harmonization algorithms do not explicitly consider downstream application performance during harmonization. However, the choice of downstream application might influence what might be considered as study-specific confounds. Therefore, ignoring downstream applications during harmonization might potentially limit downstream performance. Here we propose a goal-specific harmonization framework that utilizes downstream application performance to regularize the harmonization procedure. Our framework can be integrated with a wide variety of harmonization models based on deep neural networks, such as the recently proposed conditional variational autoencoder (cVAE) harmonization model. Three datasets from three different continents with a total of 2787 participants and 10085 anatomical T1 scans were used for evaluation. We found that cVAE removed more dataset differences than the widely used ComBat model, but at the expense of removing desirable biological information as measured by downstream prediction of mini mental state examination (MMSE) scores and clinical diagnoses. On the other hand, our goal-specific cVAE (gcVAE) was able to remove as much dataset differences as cVAE, while improving downstream cross-sectional prediction of MMSE scores and clinical diagnoses.
Introduction
In neuroimaging, "mega-analyses" combine MRI data from multiple sites to increase statistical power. However, different scanners and software pipelines introduce "site effects"—non-biological variations that can mask real clinical findings. Existing harmonization tools like ComBat (a linear model) or standard cVAEs (deep learning models) aim to make data from different sites look identical. The problem is that they often fail to distinguish between "noise" and "signal." For instance, if a model aims to harmonize data across different sites to study Alzheimer's, it might accidentally "harmonize away" relevant biological differences that are actually relevant to the disease. This paper argues that harmonization should be "goal-specific," removing unwanted site variance while explicitly protecting the information needed for a specific clinical goal.
Methods
The proposed framework, gcVAE (goal-specific conditional Variational Autoencoder), builds upon the standard cVAE architecture by incorporating downstream task performance into the training loop.
- The Baseline (cVAE): An encoder compresses brain ROI (Region of Interest) volumes into a site-invariant code. A decoder then reconstructs the original data using this code and a "site label." A discriminator ensures the encoder does not retain site-specific information in the latent code.
- The Innovation (gcVAE): The authors integrated a goal-specific Deep Neural Network (goalDNN). This goalDNN is pre-trained on the ADNI dataset to predict MMSE scores (cognitive impairment) and clinical diagnoses (e.g., Healthy vs. Alzheimer’s).
- Regularization: During training, the model is penalized if the harmonized data leads to poor predictions by the goalDNN. This forces the gcVAE to preserve biological features critical for clinical diagnosis, such as hippocampal volume, while still removing site-specific noise.
Datasets and Matching
The study utilized 10,085 anatomical T1 scans from 2,787 participants across three datasets: ADNI (North America), AIBL (Australia), and MACC (Singapore). To isolate scanner-related differences from biological ones, the researchers used a Hungarian matching algorithm to create pairs of participants across datasets who were identical in age, sex, and clinical status.
Key Results
- Site Difference Removal: Before harmonization, classifiers could identify the source dataset with 100% accuracy. Post-harmonization with gcVAE or cVAE, this dropped to ~60% (where 50% is random chance), significantly outperforming ComBat (~63-72%).
- Preserving Clinical Signal: gcVAE significantly outperformed standard cVAE in preserving diagnostic utility. In the AIBL dataset, standard cVAE dropped clinical diagnosis accuracy to 0.26, whereas gcVAE achieved 0.69—surpassing even the unharmonized data (0.48).
- MMSE Prediction: gcVAE achieved the lowest Mean Absolute Error (MAE) for cognitive scores across test datasets, validating that goal-specific constraints successfully maintain relevant biological signals.
Discussion and Limitations
While gcVAE excels at preserving signals for its specific goal (e.g., Alzheimer’s diagnosis), it is specialized. Harmonized data produced for one goal might not be optimal for another, such as studying schizophrenia or general brain aging. Future work may explore multi-goal regularization or applying this framework to raw voxel-wise image data rather than ROI volumes.
Summary updated