Modeling Misaligned Brain Images with Additive Neural Network Gaussian Processes and Regional Mixture Pooling
Abstract
Understanding relationships among brain images at different spatial resolutions—often misaligned due to hierarchical parcellation—is a key challenge in neuroimaging. We propose a non-linear regression framework to model structural response images from both structural and network-based predictors, accounting for misalignment via additive neural network Gaussian processes (ANN-GPs). This approach combines the flexibility of deep neural networks with the uncertainty quantification of Gaussian processes. To further enhance robustness and leverage limited sample sizes, we introduce a mixture modeling strategy that adaptively pools information across regions. This enables better capture of global patterns like symmetry and improves predictive inference. Simulations show our method outperforms existing techniques in modeling complex cross-modal brain image relationships.