Deep learning-based simultaneous registration and unsupervised non-correspondence segmentation of medical images with pathologies

• DOI: 10.1007/s11548-022-02577-4
• Publication Date: 2022-03-03T16:02:54Z
• Authors: Julia Andresen; Timo Kepp; Jan Ehrhardt; Claus von der Bur...; Johann Roider; Heinz Handels
• Keywords: Image registration; Robustness

The registration of medical images often suffers from missing correspondences due to inter-patient variations, pathologies and their progression leading implausible deformations that cause misregistrations might eliminate valuable information. Detecting non-corresponding regions simultaneously with the process helps generating better has been investigated thoroughly classical iterative frameworks but rarely deep learning-based methods.We present joint non-correspondence segmentation image network (NCR-Net), a convolutional neural (CNN) trained on Mumford-Shah-like functional, transferring approach field learning. NCR-Net consists one encoding two decoding parts allowing generate diffeomorphic segment non-correspondences. loss function is composed masked distance measure regularization deformation output. Additionally, anatomical labels are used for weak supervision task. No manual segmentations non-correspondences required.The proposed evaluated publicly available LPBA40 dataset artificially added stroke lesions longitudinal optical coherence tomography (OCT) patients age-related macular degeneration. data quantitatively assess performance network, it shown qualitatively can be unsupervised in OCT images. Furthermore, compared registration-only state-of-the-art algorithms showing achieves competitive superior robustness non-correspondences.NCR-Net, CNN simultaneous segmentation, presented. Experimental results show network's ability an manner its robust even presence large pathologies.