BIARD: Bayesian Individualized Abnormal Region Detection

Identifying individualized abnormal regions is a central goal in biomedical imaging, as it can provide critical insights for personalized treatment. In practice, individualized abnormal region detection poses substantial challenges, including pathological heterogeneity, shared biological variation, imaging spatial dependence, and sparsity in both biological and disease effects. To address these challenges, we propose a novel framework, Bayesian Individualized Abnormal Region Detection (BIARD), that formulates individualized abnormal region identification as a Bayesian latent-space variable selection problem. This unified Bayesian modeling strategy simultaneously accommodates heterogeneity, spatial correlation, and sparsity, enabling reliable detection of subject-specific abnormal regions in high-dimensional imaging settings. Posterior inference is conducted using an efficient Gibbs sampling algorithm, facilitating robust estimation of both model parameters and hyperparameters. Comprehensive simulation studies and a real example from the Alzheimer’s Disease Neuroimaging Initiative study demonstrate that the proposed BIARD framework substantially outperforms popular unsupervised learning and clustering-based approaches in recovering individualized abnormal regions.