Abstract
Harmonization is essential in multi-center medical imaging studies to mitigate non-biological variations known as "scanner effects." This paper introduces ComBat-MLE, a novel formulation of the gold-standard ComBat method using Maximum Likelihood Estimation (MLE). By introducing specific parameter constraints, we derive a closed-form solution that is both simpler and faster than traditional Empirical Bayes approaches. Our experiments on synthetic and phantom datasets demonstrate that ComBat-MLE effectively removes scanner noise while preserving biological signals. Furthermore, this perspective reveals ComBat as a special case of a Gaussian Mixture Model, providing a pathway for future harmonization methods that account for multiple tissue types.
The Challenge of Multi-Center Imaging
In the era of large-scale medical imaging research, data is frequently aggregated from multiple institutions. While this increases statistical power, it introduces significant "scanner effects"—variations in image features caused by differences in hardware, software versions, and acquisition protocols. These non-biological variations can obscure true clinical findings and lead to poor generalization of machine learning models across different sites.
The ComBat (Combining Batches) method, originally developed for genomics, has become the standard for removing these effects in radiomics. However, standard ComBat relies on Empirical Bayes (EB) estimation, which involves "borrowing information" across features. While effective for high-dimensional data with small samples, EB can be computationally intensive and mathematically opaque. Moreover, standard ComBat assumes a uniform transformation for all tissue types, which may not hold true in complex clinical scenarios involving both healthy and pathological tissues.
Methodology: Re-imagining ComBat via MLE
To address these limitations, we re-formulated the ComBat model as a parameter estimation problem for a multivariate Gaussian distribution. We assume that "clear" (clean) data follows a Gaussian distribution $N(\mu, \Sigma)$. The observed data, contaminated by scanner noise, is modeled as:
$$y_{is} = \mu + \gamma_s + \delta_s \epsilon_{is}$$
where $\gamma_s$ and $\delta_s$ represent additive and multiplicative scanner effects for scanner $s$, and $\epsilon_{is} \sim N(0, \Sigma)$.
Instead of the traditional EB approach, we employed Maximum Likelihood Estimation (MLE). To ensure the model parameters were identifiable, we explored two sets of constraints:
- Scanner-centric constraints: Constraints placed on the scanner effects ($\gamma_s$ and $\Delta_s$). This approach required an iterative Newton-method solver.
- Signal-centric constraints: Constraints placed on the clear data parameters ($\mu$ and $\Sigma$).
We discovered that the signal-centric approach leads to a closed-form solution that reaches its local maxima in just one iteration. This faster, more direct solution was selected as the proposed ComBat-MLE method.
Experimental Evaluation
We validated ComBat-MLE using both synthetic and physical phantom data:
Synthetic "Toy" Dataset
Using scikit-learn, we generated synthetic data simulating five different scanner settings and three distinct biological classes. This allowed us to test the method in a controlled environment where the ground truth was known.
MRI Phantom Data
We utilized two physical phantoms: one homogeneous (mimicking fluids) and one heterogeneous (mimicking white matter). These were scanned under 10 different configurations, varying the Field of View (FOV), matrix size, and scanner models. We extracted 92 radiomic features from these scans. To support the research community, we have made this feature dataset publicly available on GitHub to help other researchers test harmonization methods.
Key Results and Insights
Our experiments yielded several critical findings:
- Effective Noise Removal: In the homogeneous phantom, 78.73% of features exhibited significant scanner effects before processing. After applying ComBat-MLE, this was reduced to just 0.25%.
- Biological Preservation: We used Linear Discriminant Analysis (LDA) to classify phantom "tissues." The classification accuracy improved from 70.78% on raw data to 88.38% after ComBat-MLE harmonization, demonstrating that the method preserves essential biological information.
- Competitive Performance: ComBat-MLE performed comparably to the "ComBat-MoM" (Method of Moments) variant. Notably, it outperformed standard parametric and non-parametric EB versions in scenarios where the "information borrowing" assumption was violated.
ComBat as a Gaussian Mixture Model
A significant theoretical insight from this work is the realization that ComBat can be viewed as a special case of a Gaussian Mixture Model (GMM). This provides a clear mathematical roadmap for developing next-generation harmonization methods. Specifically, a GMM-based approach could handle multiple "pattern classes" (e.g., tumor vs. healthy tissue) simultaneously, allowing for class-specific scanner effect corrections.
Conclusion and Future Work
While ComBat-MLE does not drastically outperform existing variants in terms of raw accuracy, its value lies in its mathematical clarity, computational efficiency, and the theoretical bridge it builds to GMMs.
Future research will focus on implementing the GMM-based harmonization framework introduced in this paper. This would allow the model to learn different scanner effects for different tissue types, potentially solving complex harmonization cases where current global methods fall short.
Summary updated