Predictive Modeling for Regularization Estimation in Image Registration
Abstract:
Deformable image registration has been widely used in various medical image analysis tasks, e.g., anatomical shape analysis, atlas-based image segmentation, and motion correction in dynamic imaging. However, parameter estimation of traditional registration algorithms are oftentimes labor- consuming and computationally expensive due to the high dimensionality of medical image volumes (e.g., a brain MRI scan with the size of 2563). In this work, we introduce a predictive model for automatically estimating regularization parameters of image registration without parameter tuning. Specifically, we present a learning-based model that utilizes a deep convolutional neural network (CNN) to learn the mapping between pairwise images and the regularization parameters, which are estimated from maximum a posterior (MAP) procedure. We demonstrate the proposed methods in 2D synthetic and 3D pairwise image registration of brain studies. Our developed tool substantially improves the efficiency of image registration algorithm in various clinical scenarios, e.g. image- guided real-time neurosurgery navigation system.
Committee:
- Tom Fletcher (Chair)
- Miaomiao Zhang (Advisor)
- Yangfeng Ji
- Frederick Epstein