Mitigating Spurious Correlations in Preclinical Alzheimer's Disease Detection via Pretraining

Deep learning models suffer from shortcut learning when facing explicit spurious correlations between diagnosis labels and non-neuropathological information (e.g., gender, age, clinical protocol). However, the effects of implicit spurious correlations caused by implicit non-neuropathological information are difficult to be analyzed statistically. In this paper, we investigate shortcut learning of implicit spurious correlations in sMRI-based preclinical AD detection with classification performance and neuropathological validation. We evaluate two AD-related supervised pretraining tasks (i.e., AD detection, multiclassification) and three self-supervised pretraining tasks (i.e., auto-encoding, denoising, mask reconstruction) on removal of implicit spurious correlations. In contrast to earlier works in other performance metrics, we do not observe that self-supervised pretraining tasks achieve a significant performance improvement in mitigating implicit spurious correlation. Instead, we find that AD-related supervised tasks, especially multiclassification, are effective for deep models to model the severity of disease and capture neuropathological features. Moreover, we conduct the same experiment on another dataset in mild cognitive impairment detection. The results validate our conclusions.