2021/12/08 by Alessa Hering, Hering, Alessa, Lasse Hansen +97 · 23 citations
Engineering · Medicine · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image and Video Processing (eess.IV) #Medical Imaging Techniques and Applications #Medical Imaging and Analysis #Radiomics and Machine Learning in Medical Imaging #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2112.04489
openalex publication_date 2021/12/08 · openalex created_date 2022/05/05 · openalex updated_date 2026/07/28
Image registration is a fundamental medical image analysis task, and a wide\nvariety of approaches have been proposed. However, only a few studies have\ncomprehensively compared medical image registration approaches on a wide range\nof clinically relevant tasks. This limits the development of registration\nmethods, the adoption of research advances into practice, and a fair benchmark\nacross competing approaches. The Learn2Reg challenge addresses these\nlimitations by providing a multi-task medical image registration data set for\ncomprehensive characterisation of deformable registration algorithms. A\ncontinuous evaluation will be possible at\nhttps://learn2reg.grand-challenge.org. Learn2Reg covers a wide range of\nanatomies (brain, abdomen, and thorax), modalities (ultrasound, CT, MR),\navailability of annotations, as well as intra- and inter-patient registration\nevaluation. We established an easily accessible framework for training and\nvalidation of 3D registration methods, which enabled the compilation of results\nof over 65 individual method submissions from more than 20 unique teams. We\nused a complementary set of metrics, including robustness, accuracy,\nplausibility, and runtime, enabling unique insight into the current\nstate-of-the-art of medical image registration. This paper describes datasets,\ntasks, evaluation methods and results of the challenge, as well as results of\nfurther analysis of transferability to new datasets, the importance of label\nsupervision, and resulting bias. While no single approach worked best across\nall tasks, many methodological aspects could be identified that push the\nperformance of medical image registration to new state-of-the-art performance.\nFurthermore, we demystified the common belief that conventional registration\nmethods have to be much slower than deep-learning-based methods.\n