Motivation: Deep learning models in MRI are prone to learn spurious correlations, induced by confoundings of e.g. scanners, patient demographics, instead of operating on the true causal dependencies. This can lead to reduced generalization and biased predictions. Goal(s): We aim to develop a robust framework that eliminates multiple spurious correlations, enabling predictions based on causal relationships. Approach: Our novel framework, MIMM-X, leverages a Confounder-Attention-Summarizer to mitigate multiple spurious correlations. It disentangles casual features from spuriously correlated features and minimizes the mutual information between them. Results: MIMM-X demonstrates improved generalization and robustness in diverse MRI settings, effectively removing spurious correlations across scanners and patient populations. Impact: Our novel framework, MIMM-X, an extension of our previous model (MIMM), is able to remove multiple spurious correlations in MRI, ensuring causal predictions based on task-relevant features. This approach improves generalization for data across various MR scanners and patient demographics.
Fay et al. (Tue,) studied this question.