Designing implicit population learners: a permutation-equivariant state space approach for brain disease diagnosis

Area of Science:

• Neuroimaging Analysis
• Machine Learning
• Computational Neuroscience

Background:

• Group-aware learning enhances neuroimaging disease diagnosis by leveraging population interactions.
• Existing graph-based methods face scalability and design limitations due to explicit graph construction.

Purpose of the Study:

• Introduce IP-Mamba, a scalable and memory-efficient framework for implicit population interaction modeling in neuroimaging.
• Address limitations of explicit graph construction in group-aware learning for disease diagnosis.

Main Methods:

• Utilize a bidirectional Mamba-based sequence modeling approach on unordered subject sets.
• Implement a Shuffle Consistency Strategy to ensure permutation equivariance and set-based modeling.
• Employ Contextual Population Support Set inference and a hybrid SVM for classification, addressing class imbalance.

Main Results:

• Achieved 87.84% balanced accuracy and 89% sensitivity for the minority class in Alzheimer's disease classification (OASIS-1 dataset).
• Demonstrated linear O(N) memory scaling, avoiding quadratic bottlenecks of graph-based attention networks.
• IP-Mamba showed competitive diagnostic robustness compared to 3D CNNs and Transformers.

Conclusions:

• IP-Mamba provides a principled, memory-efficient, and scalable alternative to explicit graph-based methods for population-aware neuroimaging analysis.
• The framework is particularly effective in imbalanced clinical settings, offering robust disease detection.
• Bidirectional modeling and shuffle consistency are crucial components for IP-Mamba's performance.