A lightweight generative model for interpretable subject-level prediction.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Discriminative models accurately predict diagnoses from medical images but lack anatomical interpretability.
• Classical human brain mapping techniques use generative models to encode cause-effect relations.

Purpose of the Study:

• To develop a simple, inherently interpretable technique for single-subject prediction from medical images.
• To augment generative models with a noise model for enhanced prediction and explanation.

Main Methods:

• Augmenting generative models with a multivariate noise model to capture spatial correlations.
• Efficient inversion of the model for subject-level predictions.
• Incorporating cause-effect encoding from classical brain mapping.

Main Results:

• The proposed method achieves accurate subject-level predictions.
• The model provides intuitive visual explanations of its predictions.
• The technique is efficient, with fast training and a single hyperparameter.

Conclusions:

• The developed method offers a powerful tool for interpretable medical image analysis.
• It bridges the gap between predictive accuracy and anatomical understanding in neuroimaging.
• The approach facilitates easier adoption due to its simplicity and speed.