Our goal is to develop a subject-specific interpretable deep learning-based framework. The project will be structured as follows:
Aim 1. To develop a deep learning-based regression model for biological age predictions at both global and regional levels: Brain age serves as a personalized biomarker for assessing brain health, reflecting deviations from typical neuroanatomic aging trajectories within individuals. However, brain regions exhibit varying rates of aging in different brain regions. Therefore, this task seeks to develop a brain age regression model capable of predicting the age of an individual’s brain not only at a global level but also for each specific anatomical region by using T1 structural MRI datasets from multiple open-source databases.
Aim 2. To develop a subject-specific XAI framework for risk factor identification: Understanding the important features contributing to brain age predictions using deep learning models is essential for clinical implementation. Here, we will utilize Shapley value analysis to fairly attribute the prediction score to feature values for contrastive explanation. This local and global brain age explanation will reveal the image based features for accelerated biological aging, with potential implications for early biomarkers in AD patients. These features will also be used to identify bias due to statistical imparity in the dataset especially for women and racial/ethnic minorities.
