Skull-stripping induces shortcut learning in MRI-based Alzheimer's disease classification

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep neural networks achieve high accuracy in classifying Alzheimer's disease (AD) from structural MRI.
• The specific image features driving these classifications remain unclear, posing challenges for clinical trust.

Purpose of the Study:

• To systematically assess the contributions of T1-weighted (T1w) MRI texture, volumetric information, and preprocessing steps, particularly skull-stripping, to AD classification.
• To investigate the mechanisms underlying deep learning-based disease classification in AD.

Main Methods:

• Utilized a dataset of 990 T1w MRIs from AD patients and controls (ADNI database).
• Varied preprocessing through skull-stripping and intensity binarization to isolate texture and shape contributions.
• Trained 3D convolutional neural networks and analyzed feature relevance using Layer-wise Relevance Propagation and spectral clustering.

Main Results:

• Classification accuracy, sensitivity, and specificity remained stable across different preprocessing conditions.
• Models showed minimal reliance on gray-white matter texture, performing similarly on binarized images.
• Volumetric features, specifically brain contours introduced by skull-stripping, were consistently utilized by the models.

Conclusions:

• Deep learning models exhibited shortcut learning, relying on preprocessing artifacts (skull-stripping) rather than genuine disease-related features.
• This underscores the critical importance of interpretability tools to detect biases and ensure robust, trustworthy AI in medical imaging.
• Careful consideration of dataset characteristics and preprocessing is essential for developing clinically relevant and validated deep learning models.