Weakly-supervised AI methods dissect the neuroanatomical heterogeneity of aging and brain diseases

Disease heterogeneity remains a major obstacle to achieving precision diagnostics. Artificial intelligence (AI) advances have introduced promising approaches to tackle this complexity, particularly by identifying imaging-derived biomarkers capable of predicting disease progression and mortality. Moving beyond traditional unsupervised clustering methods like K-means, Dr. Davatzikos and his team have developed a novel weakly supervised learning framework. This approach leverages generative adversarial networks (GANs) and non-negative matrix factorization methods, in addition to adversarial autoencoders, to model disease trajectories from the healthy control domain to the patient domain, capturing biologically meaningful variance while reducing confounding influences such as demographic factors. In this talk, Dr. Davatzikos will introduce several weakly-supervised AI methods to unravel the neuroanatomical heterogeneity of aging and brain diseases. Among these are Smile-GAN (PMID: 34862382), a model that identifies imaging-derived disease subtypes, and Surreal-GAN (PMID: 39147830), which provides a continuous representation of disease heterogeneity through representation learning. He will also discuss Gene-SGAN (PMID: 38191573), a method that builds upon Smile-GAN by integrating imaging and genetic data to enhance understanding of disease heterogeneity. Beyond methodological innovation, Dr. Davatzikos will highlight the application of these models in diverse contexts, including normal aging and disorders such as Alzheimer’s disease and schizophrenia. This talk will exemplify how advanced AI and imaging techniques can model disease heterogeneity to advance precision medicine.