Tooth instance segmentation from cone-beam CT images through point-based detection and Gaussian disentanglement

Individual tooth segmentation and identification from cone-beam computed tomography images are preoperative prerequisites for orthodontic treatments. Instance methods using convolutional neural networks have demonstrated ground-breaking results on individual tasks, used in various medical imaging applications. While point-based detection achieve superior dental images, it is still a challenging task to distinguish adjacent teeth because of their similar topologies proximate nature. In this study, we propose localization network that effectively disentangles each based Gaussian disentanglement objective function. The proposed first performs heatmap regression accompanied by box all the anatomical teeth. A novel penalty employed minimizing sum pixel-wise multiplication heatmaps pairs. Subsequently, performed converting labeling distance map minimize false positives regions Experimental demonstrate algorithm outperforms state-of-the-art approaches increasing average precision 9.1%, which high performance terms segmentation. primary significance method two-fold: (1) introduction framework does not require additional classification (2) design loss function separates distributions responses framework.