2019/07/25 by Mattias P. Heinrich, Heinrich, Mattias P. · 5 citations
Computer Science · Engineering · Medicine · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Medical Image Segmentation Techniques #Medical Imaging Techniques and Applications #Medical Imaging and Analysis
paper · pdf · doi:10.48550/arxiv.1907.10931
openalex publication_date 2019/07/25 · openalex created_date 2022/07/19 · openalex updated_date 2026/07/28
Nonlinear image registration continues to be a fundamentally important tool\nin medical image analysis. Diagnostic tasks, image-guided surgery and\nradiotherapy as well as motion analysis all rely heavily on accurate\nintra-patient alignment. Furthermore, inter-patient registration enables\natlas-based segmentation or landmark localisation and shape analysis. When\nlabelled scans are scarce and anatomical differences large, conventional\nregistration has often remained superior to deep learning methods that have so\nfar mainly dealt with relatively small or low-complexity deformations. We\naddress this shortcoming by leveraging ideas from probabilistic dense\ndisplacement optimisation that has excelled in many registration tasks with\nlarge deformations. We propose to design a network with approximate\nmin-convolutions and mean field inference for differentiable displacement\nregularisation within a discrete weakly-supervised registration setting. By\nemploying these meaningful and theoretically proven constraints, our learnable\nregistration algorithm contains very few trainable weights (primarily for\nfeature extraction) and is easier to train with few labelled scans. It is very\nfast in training and inference and achieves state-of-the-art accuracies for the\nchallenging inter-patient registration of abdominal CT outperforming previous\ndeep learning approaches by 15% Dice overlap.\n