Interpretable brain disease classification and relevance-guided deep learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep neural networks (DNNs) are used for neurological disease classification using MRI.
• Interpretability of DNN decisions in medical imaging remains a challenge.
• Existing methods may rely on non-physiological features outside brain tissue.

Purpose of the Study:

• To develop a regularization technique for training interpretable convolutional neural network (CNN) classifiers for Alzheimer's disease.
• To improve classification accuracy and focus on relevant brain regions.
• To mitigate preprocessing artifacts in MRI analysis.

Main Methods:

• Utilized deep Taylor decomposition for heat mapping to identify crucial image features.
• Developed a relevance-guided regularization technique for CNN training.
• Applied the method to T1-weighted MRI from Alzheimer's disease patients and control subjects.

Main Results:

• The relevance-guided framework achieved higher classification accuracies compared to conventional CNNs.
• The model focused on fewer, more relevant, and physiologically plausible voxels within brain tissue.
• Preprocessing effects from skull stripping and registration were mitigated.

Conclusions:

• The developed framework enhances the interpretability of CNNs in Alzheimer's disease classification.
• Relevance-guided CNNs offer a more physiologically grounded approach to MRI-based diagnosis.
• This challenges the reliance on unprocessed MRI data for high classification accuracy.