2D/3D image registration using deep learning

• Main Author: Dai, Yuhe
• Other Authors: Huang Weimin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/167753
• Institution: Nanyang Technological University

Image registration is a fundamental task in computer vision, particularly in medical image processing. It has numerous applications in the medical field, but medical images often contain undesirable artifacts that can compromise the accuracy of registration results. Additionally, current CNN-based methods may not fully consider the long-range spatial relationships within images, which can limit their performance. Thus, this project proposes a two-stage method for medical image registration that corrects bias and registers images with intensity inhomogeneity. The first stage introduces a robust non-convex regularizer to recover the true intensity, and a smooth regularizer to model the bias field, along with a data fidelity term that incorporates local intensity characteristics. In the second stage, a state-of-art unsupervised DIR model called TransMorph is utilized, which is built on Transformer and capable of establishing spatial correspondences between image voxels over long distances. The method is tested on a T1-weighted brain MRI dataset from CMI-HBN and compared with an existing method. It finally achieves a more robust and accurate registration result, confirming the effectiveness of this two-stage method for medical image registration.