Avoiding Shortcut-Learning by Mutual Information Minimization in Deep Learning-Based MR Image Processing

Deep Learning methods can detect patterns in data such as MR images but are incapable of determining causal relationships. However, causal understanding is crucial in medical applications, since the presence of confounders (e.g. scan conditions) obscure the causal relationship and create spurious-correlations. State-of-the-art models purely rely on correlated patterns which can result in wrong conclusions or diagnoses when spurious-correlations change (e.g. new scanner). We propose a deep learning framework that is robust in the presence of spurious-correlations by decreasing mutual information between learned features of MR images and leads to improved performance under distribution shifts.