TL;DR
This paper introduces a "fully Bayesian" version of ComBat (FB ComBat) for harmonizing neuroimaging data, aiming to reduce scanner-induced biases while preserving biological information. FB ComBat offers richer uncertainty estimates than the standard empirical Bayes (EB) ComBat, leading to better performance in tasks like disease prediction and more principled statistical analyses, though it is computationally more intensive.
What Problem Does This Paper Address?
Combining neuroimaging data from multiple sources (different scanners, acquisition protocols, etc.) is crucial for large-scale studies, but it introduces unwanted variability (bias) that can obscure true biological signals. Harmonization techniques aim to remove this scanner-related bias. ComBat is a popular method, but its standard implementation using empirical Bayes (EB ComBat) tends to underestimate uncertainty. This paper addresses the need for a harmonization method that not only removes bias but also provides more accurate uncertainty estimates, which are vital for downstream analyses like disease prediction and understanding biological variability.
Methods
The core contribution is the development of a Fully Bayesian ComBat (FB ComBat). Unlike EB ComBat, which uses point estimates for hyperparameters, FB ComBat infers full posterior distributions for all model parameters using Markov Chain Monte Carlo (MCMC) sampling, specifically Hamiltonian Monte Carlo (HMC) with a No-U-Turn Sampler (NUTS). This allows for a more principled estimation of uncertainty. The authors also compare FB ComBat to the standard EB ComBat and an unharmonized baseline across various evaluation metrics.
Datasets
The primary dataset used is from the Alzheimer's Disease Neuroimaging Initiative (ADNI). This dataset includes T1-weighted structural MRI images from 809 subjects with varying cognitive statuses (cognitively normal, mild cognitive impairment, Alzheimer's disease), acquired on 83 different scanners.
Key Insights and Ideas
- Uncertainty Matters: The paper strongly advocates for the importance of accurate uncertainty estimation in harmonization. FB ComBat's richer posterior distributions provide a more realistic picture of this uncertainty, which is crucial for downstream tasks.
- Biological vs. Scanner Information: There's a trade-off between removing scanner bias and preserving biological information (like age and disease effects). FB ComBat appears to strike a better balance, retaining more biological relevance than EB ComBat in some aspects.
- Generative Power of Posteriors: The posterior distributions from FB ComBat can be used for data augmentation, effectively creating synthetic data points that improve classifier performance, especially in limited-data scenarios.
- Region-Specific Uncertainty: Not all brain regions are equally susceptible to measurement uncertainty. Subcortical volumes tend to be more reliable than cortical thickness, with specific regions like the pericalcarine cortex showing higher uncertainty.
Key Results
- Scanner Effect Removal: Both EB ComBat and FB ComBat significantly reduce scanner effects. FB ComBat showed slightly better performance in removing bias and variance in test-retest scans.
- Biological Information Retention: FB ComBat demonstrated stronger performance in age and AD classification tasks compared to EB ComBat, suggesting better retention of biologically relevant information.
- Data Augmentation: Using FB ComBat harmonization with posterior distribution resampling for augmentation led to the highest Area Under the Receiver Operating Characteristic Curve (AUROC) for MCI vs. AD classification (0.821), outperforming EB ComBat and unharmonized data.
- Computational Cost: FB ComBat is significantly slower than EB ComBat due to MCMC sampling, taking about six hours for their dataset compared to EB ComBat's efficiency.
- Scanner Strength Prediction: EB ComBat was slightly better at removing scanner strength information (lower AUROC for predicting scanner strength), indicating it might be more effective at removing non-biological scanner effects in this specific task.
Limitations and Future Work
- Computational Cost: The primary limitation of FB ComBat is its computational intensity compared to EB ComBat.
- Ground Truth: Evaluating harmonization is inherently difficult due to the unknown ground truth of true biological measurements. Test-retest experiments and simulations serve as proxies.
- Model Complexity: The authors suggest that more complex FB models, potentially incorporating non-linear covariate effects or scanner covariance, should be explored and compared to their EB ComBat counterparts.
- Type I Error Mitigation: Further research is needed to develop strategies for mitigating Type I error in harmonization, particularly for additive and multiplicative scanner error terms.
- Generalizability: While tested on ADNI, the performance of FB ComBat on other imaging modalities (PET, white matter) and different disease contexts warrants further investigation.
Summary updated