SCALABLE MATCHING OF DEFORMED IMAGES

Image registration is a fundamental inverse problem ubiquitous across biomedical and life science applications. Over the past three decades, advances in imaging technology have democratized access to unprecedented spatial and temporal detail, while an equally staggering growth in GPU hardware has transformed high-performance computing domains including machine learning, computational fluid dynamics, and molecular dynamics. Despite methodological progress in image registration, algorithms have not scaled in tandem with these advances, constraining researchers to work with heavily downsampled versions of the rich data they acquire. This dissertation presents a comprehensive investigation into a Pareto-efficient image registration framework that simultaneously optimizes for accuracy, robustness, and scalability. First, we conduct a systematic empirical study comparing classical optimization-based and deep learning paradigms, addressing instrumentation bias in the literature and providing practitioners with a principled framework for selecting appropriate methods. Second, we introduce FireANTs, a GPU-accelerated framework for adaptive Riemannian optimization on the space of diffeomorphisms. We formally quantify the ill-conditioning of deformable registration and develop a novel Eulerian descent formulation enabling powerful adaptive optimization on time-dependent diffeomorphic flows, achieving state-of-the-art zero-shot performance across multiple imaging modalities and benchmark datasets with orders-of-magnitude speedups over existing methods. Third, we develop Deep Implicit Optimization (DIO), which transforms FireANTs into a fully differentiable layer within deep networks. By decoupling feature learning from optimization, DIO inherits task-specific appearance invariances from the optimization objective while enabling end-to-end learning of dense multi-scale features from weak supervision signals, yielding superior generalization to domain shifts and zero-cost plug-and-play of arbitrary transformation representations at test time. Fourth, we develop system-level innovations to scale registration to gigavoxel imaging problems, including IO-aware fused CUDA kernels that reduce memory overheads for bottleneck operations, enabling multimodal registration on previously intractable datasets in minutes. Finally, we demonstrate real-world impact through applications including multimodal in-vivo brain to ex-vivo hemisphere registration, geometry distortion correction for echo-planar MRI, and gradient-free landmark-guided registration for lung CT scans.

Author dc:creator

• Jena, Rohit, Kumar

Advisors dc:contributor.advisor

• Gee, James, C
• Chaudhari, Pratik

• Medical Sciences
• Computer Sciences

Language dc:language.iso

en

Repository record dc:identifier.uri

https://repository.upenn.edu/handle/20.500.14332/62749

OAI identifier oai:identifier

oai:repository.upenn.edu:20.500.14332/62749