Bingfan Liu
Title: Dynamic Survival Prediction using Longitudinal Images
Date: August 8th, 2025
Time: 2:00pm
Location: LIB 7200
Supervised by: Jiguo Cao & Haolun Shi
Abstract:
Advancements in imaging and information technology lead to the widespread collection of longitudinal high-dimensional medical images, offering valuable prognostic insights in clinical studies. Accurate prognosis for progressive illnesses often necessitates analysis of the entire sequence of medical images. However, the unstructured nature of image data, combined with its temporal correlations and high dimensionality, presents significant challenges for survival prediction. This thesis introduces three novel methods to address these obstacles.
Furthermore, the expansion of Internet and storage technologies enables access to largescale medical datasets. While, efficient analysis of such vast datasets is essential, it is often tampered by computational constraints, including memory limitations and processing time. To overcome these challenges, traditional approaches rely on parallel and distributed computing techniques. In this thesis, we propose an innovative subsampling method tailored for generalized additive models. Our method enhances computational efficiency while preserving the flexible modeling capabilities, offering a solution to big data challenges.
Keywords: functional data analysis; longitudinal image; mamba; optimal sampling; survival analysis; transformer