CNN-based lung CT registration with multiple anatomical constraints

Deep-learning-based registration methods emerged as a fast alternative to conventional methods. However, these often still cannot achieve the same performance because they are either limited small deformation or fail handle superposition of large and deformations without producing implausible fields with foldings inside. In this paper, we identify important strategies for lung successfully developed deep-learning counterpart. We employ Gaussian-pyramid-based multilevel framework that can solve image optimization in coarse-to-fine fashion. Furthermore, prevent field restrict determinant Jacobian physiologically meaningful values by combining volume change penalty curvature regularizer loss function. Keypoint correspondences integrated focus on alignment smaller structures. perform an extensive evaluation assess accuracy, robustness, plausibility estimated fields, transferability our approach. show it achieves state-of-the-art results COPDGene dataset compared method much shorter execution time. experiments DIRLab exhale inhale registration, demonstrate substantial improvements (TRE below 1.2 mm) over other deep learning Our algorithm is publicly available at https://grand-challenge.org/algorithms/deep-learning-based-ct-lung-registration/.