Confounder-Aware Visualization of ConvNets.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnosis

Background:

• Deep learning, particularly convolutional networks (ConvNets), is widely used in neuroimaging for disease diagnosis from MR images.
• Current visualization techniques (saliency maps) often highlight brain regions influenced by confounding variables (e.g., age) rather than the diagnosis itself, leading to misinterpretation.

Purpose of the Study:

• To develop a method for generating confounder-free saliency maps in ConvNet-based neuroimaging analysis.
• To ensure that visualized brain regions are solely predictive of the diagnosis, improving model interpretability.

Main Methods:

• Incorporation of univariate statistical tests to identify confounding effects in intermediate ConvNet features.
• A novel partial back-propagation technique to remove the influence of confounded features.
• Application of the approach to synthetic and real neuroimaging datasets.

Main Results:

• The proposed method successfully generated confounder-free saliency maps.
• Visualizations accurately highlighted brain voxels predictive of diagnosis, free from confounding influences.
• Demonstrated potential for unbiased interpretation of deep learning models in neuroimaging.

Conclusions:

• The developed approach effectively addresses the confounding issue in ConvNet saliency maps for neuroimaging.
• This method enhances the reliability and interpretability of deep learning models used for medical diagnosis.
• Enables more accurate understanding of disease impact on the brain by providing unbiased visualizations.