Harmonization with Flow-Based Causal Inference

Rongguang Wang, Pratik Chaudhari, Christos Davatzikos

TL;DR

This paper introduces a novel method for harmonizing medical imaging data from different sources by leveraging causal inference and normalizing flows. By explicitly modeling the causal relationships between known confounders (like site, age, and sex) and imaging features, their approach generates "counterfactual" data that appears to come from a single, consistent source, leading to improved generalization for downstream tasks like age prediction and disease classification.

What Problem Does This Paper Address?

Medical imaging datasets are often collected across various hospitals, scanners, and patient cohorts, leading to significant heterogeneity. This "site effect" is a major hurdle for deep learning models, which tend to overfit to specific datasets and generalize poorly to new, unseen data. Traditional methods for "data harmonization" aim to remove these site-specific variations, but they often do so by discarding potentially useful information, including subject-specific biological variability, which can negatively impact predictive accuracy for tasks like disease diagnosis. The paper addresses the need for harmonization methods that can effectively remove unwanted site effects while preserving crucial biological information.

Methods

The core of the proposed method is a Structural Causal Model (SCM) that explicitly models the causal relationships between observed data and known confounders. This SCM is parameterized using normalizing flows, which are powerful deep learning models capable of learning complex probability distributions through a series of invertible transformations.

The process involves:

  1. Learning the SCM: A normalizing flow model is trained to represent the joint distribution of the observed data and confounders.
  2. Abduction: The model infers the posterior distribution of the exogenous noise variables given the observed data.
  3. Intervention: The value of a specific confounder (e.g., imaging site) is "intervened" upon, meaning it's replaced with a desired value. This creates a modified SCM.
  4. Counterfactual Generation: Samples are drawn from this modified SCM to generate "counterfactual" data. This data is essentially what the original observation would have looked like if it had been acquired under the intervened condition (e.g., from the target site).

This approach allows for generating harmonized data by essentially "rewriting" the data to conform to a target site's characteristics while preserving other important biological factors as exogenous noise.

Datasets

The paper evaluates its method on several large-scale, real-world medical datasets:

  • iSTAGING consortium: This dataset includes data from multiple scanners and sites (BLSA, SHIP, UKBB) and is used for age prediction and feature distribution analysis.
  • Alzheimer's Disease Neuroimaging Initiative (ADNI): This dataset is used for Alzheimer's disease classification.
  • Baltimore Longitudinal Study of Aging (BLSA): Used as part of the iSTAGING consortium.
  • Study of Health in Pomerania (SHIP): Used as part of the iSTAGING consortium.
  • UK Biobank (UKBB): Used as part of the iSTAGING consortium.

No new benchmarks were introduced; the focus was on evaluating against existing datasets and methods.

Key Insights and Ideas

  • Causality as a Harmonization Tool: Explicitly modeling causal relationships provides a principled way to disentangle unwanted site effects from genuine biological variability.
  • Preserving Biological Information: Unlike traditional methods that might remove all "unknown" confounders, this approach treats them as exogenous noise, preserving subject-specific information crucial for tasks like disease diagnosis.
  • Normalizing Flows for Counterfactuals: The invertible nature of normalizing flows is perfectly suited for generating counterfactuals by allowing for both forward (generating data) and backward (inferring noise) transformations within the causal model.
  • Site-Specific Data as "Counterfactuals": The core idea is to ask "what would this scan look like if it were acquired from site X?" and then generate that data, effectively creating a harmonized dataset.

Key Results

  • Improved Generalization: The proposed flow-based SCM method significantly outperforms state-of-the-art harmonization algorithms (like ComBat and IRM) in both age prediction and Alzheimer's disease classification tasks on target sites.
  • Reduced Mean Average Error (MAE): For age prediction, the quadratic-spline variant of the flow-based SCM achieved substantially smaller MAEs across all target sites compared to baselines. For example, on the UKBB target site, the MAE was 6.44, compared to 15.95 for ComBat.
  • Higher Classification Accuracy: In AD classification, the Q-Spline variant of the flow-based SCM achieved higher accuracies on target sites compared to ComBat and IRM. For instance, on ADNI-2 to ADNI-1 transfer, the Q-Spline achieved 77.5% accuracy, while ComBat achieved 71.1%.
  • Preservation of Feature Distributions: Visualizations (Figure 2-4) show that while ComBat methods align feature distributions across sites by removing variations, the proposed SCM method (especially Q-Spline) results in distributions that are largely unchanged from the raw data, indicating preservation of biological variability.

Limitations and Future Work

  • Known Confounders: The method relies on explicitly knowing the confounders (site, age, sex). Identifying and modeling all relevant confounders can be challenging in practice.
  • Causal Graph Identification: The paper assumes a known causal graph structure. Future work could explore methods for automatically identifying the causal graph from data.
  • Causal Mediation Analysis: The authors suggest extending the work to causal mediation analysis, which would allow understanding how confounders influence outcomes through intermediate variables.
  • Computational Cost: While not explicitly detailed as a limitation, training complex normalizing flow models can be computationally intensive.

Summary updated

Abstract

Heterogeneity in medical data, e.g., from data collected at different sites and with different protocols in a clinical study, is a fundamental hurdle for accurate prediction using machine learning models, as such models often fail to generalize well. This paper leverages a recently proposed normalizing-flow-based method to perform counterfactual inference upon a structural causal model (SCM), in order to achieve harmonization of such data. A causal model is used to model observed effects (brain magnetic resonance imaging data) that result from known confounders (site, gender and age) and exogenous noise variables. Our formulation exploits the bijection induced by flow for the purpose of harmonization. We infer the posterior of exogenous variables, intervene on observations, and draw samples from the resultant SCM to obtain counterfactuals. This approach is evaluated extensively on multiple, large, real-world medical datasets and displayed better cross-domain generalization compared to state-of-the-art algorithms. Further experiments that evaluate the quality of confounder-independent data generated by our model using regression and classification tasks are provided.
BibTeX
@inproceedings{wang2021harmonization,
  author = {Rongguang Wang and Pratik Chaudhari and Christos Davatzikos},
  title = {{Harmonization with Flow-Based Causal Inference}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=eMDzUX0kTgy},
}
Figure