Deep learning models are widely applied in medical image analysis, such as magnetic resonance imaging (MRI), to detect patterns and correlations. However, conventional models often fail to account for the underlying causal relationships in the data. In the presence of confounding factors, spurious correlations between the imaging process, image content, and labels may cause the network to learn shortcuts--resulting in biased or incorrect predictions.

This challenge becomes even more severe when applying the model to new environments or out-of-distribution (OOD) data, where these spurious correlations may not hold. As a result, such models risk generating misleading conclusions or diagnoses.

In our work (Fay et al., 2023), we introduce the Mutual Information Minimization Model (MIMM), a novel framework that enhances causal prediction while mitigating the effect of spurious correlations.

MIMM encodes the input image into a feature representation, which is then split into two disjoint components:

• One for predicting the primary task (e.g., disease classification),
• One for predicting the spuriously correlated factor (e.g., demographic attribute, scanner type).

We hypothesize that minimizing the mutual information (MI) between these two components encourages their independence--leading to confounder-free and causally meaningful predictions.

We evaluate MIMM on five datasets:

• Two non-medical benchmarks: Morpho-MNIST and Fashion-MNIST.
• Three medical imaging cohorts: German National Cohort, UK Biobank, and ADNI.

The results demonstrate that MIMM consistently outperforms conventional models by learning invariant, robust, and causally aligned representations.

1. Feature disentanglement for separate prediction of the primary task and spurious factors.
2. Mutual information minimization to prevent shortcut learning and enforce invariance to counterfactual scenarios.
3. Applicability to heterogeneous medical data through principled modeling of spurious correlations.

In standard models, confounding introduces spurious correlations that may mislead the learning process. MIMM explicitly interrupts these shortcuts and enables the model to focus on the true causal signal.

Install the required Python packages via:

To run the model with a specified configuration file:

If you use MIMM in your research, please cite our work: