A Rigid Registration Pipeline for TMJ CBCT Time-Series Images with Removal of Remodel Structures

The accurate estimation of structure and position in relation to Temporomandibular Joint (TMJ) is crucial for Temporomandibular Joint Disorder (TMD) diagnosis and treatment. However, existing automatic rigid registration methods, such as Scale Invariant Feature Transform (SIFT), often have some limitations, including reliance on artificial external features, requiring high computational effort, or ignoring the deformation and physiological changes in anatomical structures. To address these problems, we propose a pipeline for TMJ CBCT time-series images to predict the matrix of translations and rotations in rigid registration task. In our proposed pipeline, a segmentation network is employed to eliminate the effects of complex anatomical structures, and a landmark-based 3D rigid registration is adopted for preliminary rigid registration for avoid poor performance on images with large positional variations. Finally, a self-supervised neural network trained on real medial images is used to predict the rotation and translation parameters. Our network is evaluated on 44 pairs of clinical images. SIRS-Net+ achieves an average Dice coefficient of 0.8020 (approximately 8% higher than SIFT) and an average NCC of 0.7264 (about 10% higher than SIFT).