Multimodal Brain Age Estimation

Sammanfattning: Machine learning models trained on MRI brain scans of healthy subjects can be used to predict age. Accurate estimation of brain age is important for reliably detecting abnormal aging in the brain. One way to increase the accuracy of predicted brain age is through using multimodal data. Previous research using multimodal data has largely been non-deep learning-based; in this thesis, we examine a deep learning model that can effectively utilize several modalities. Three baseline models were trained. Two used T1-weighted and T2- weighted data, respectively. The third model was trained on both T1- and T2- weighted data using high-level fusion. We found that using multimodal data reduced the mean absolute error of predicted ages. Afourth model utilized disentanglement to create a representation robust to missing T1- or T2-weighted data. Our results showed that this model performed similarly to the baselines, meaning that it is robust to missing data and at no significant cost of prediction accuracy.