A one-stage deep learning method for fully automated mesiodens localization on panoramic radiographs

[Display omitted] •Developing a fully-automatic one-stage mesiodens localization approach without any manual operation.•Constructing a deep network unifying mesiodens localization and identification into one stage.•Be able to simultaneously identify and locate the mesiodens in the panoramic image.•Improving the performance of the mesiodens localization in a more cost-efficient way by replacing cross-stage partial modules in the backbone network with ghost bottlenecks and squeeze-and-excitation layers. Developing computer-aided techniques to automatically detect mesiodens on panoramic images is desirable. Most of the existing methods of mesiodens detection need to manually search and crop potential positions of the mesiodens from a new testing image before classification. As a result, they cannot automatically provide the exact location of the mesiodens in a panoramic image to identify their presence. To handle the above problems, this paper develops a fully-automatic deep learning-based approach to localize mesiodens in panoramic images at one stage. The principle behind our approach is fundamentally different from the prior ones. It treats mesiodens detection as a regression problem instead of merely an identification task. And our method constructs a deep mesiodens localization network (DMLnet) which unifies mesiodens localization and identification into one stage. Scanning the whole image only once allows it to simultaneously identify and locate the mesiodens without any manual operation. Besides, the modified backbone network makes the proposed DMLnet perform well with less computation complexity. We evaluate the performance of the proposed method on a database including primary, mixed, and permanent dentition groups. The experimental results validate the effectiveness and efficiency of our method in improving the accuracy of mesiodens detection for panoramic images. Besides, our approach can achieve competitive performance with lower computation costs than the other state-of-the-art methods.