Deformable Registration of Low-overlapping Medical Images

Abstract

Even though whole-body MRI becomes more accessible, its use is still restricted by technical limitations such as field of view and resolution. To minimize artifacts caused by respiratory motion, the acquisition time can be reduced to a feasible breath-hold by decreasing the image size. Conversely, a series of acquisitions is required to cover a larger extent. While the method is effective for individual acquisitions, different respiratory states introduce artifacts when a composite image is reconstructed from the series. In this paper, we propose a deformable registration method for low-overlapping MRI to compensate for such artifacts and facilitate seamless mosaicing. Based on an unsupervised learning-based model, our method generalizes well to different modalities and target anatomies. We demonstrate this on a dataset of 16 abdominal MRI series from a medical use case as well as synthetic image pairs generated from a large heterogeneous dataset, with 13% to 24% overlap. The evaluation shows an improved Dice Similarity Coefficient (DSC) for target structures in the overlap region by +0.14 (from 0.73) for real and +0.21 (from 0.68) for synthetic image pairs. Our method is fast and robust and may be applied to various mosaicing tasks.