Benchmark and Boosted Segmentation on Dental Panoramic Radiographs

Kaiwei Sun, YANG FENG, Zuozhu Liu

Keywords: Panoramic radiographs, Orthodontic, Segmentation, Mask2former

TL;DR: This paper provides a preliminary insight into the segmentation task of five oral structures on Panoramic radiographs. We propose a novel Cascaded Multi-scale Mask2former (CMMask2former) method for this challenging segmentation task.

Abstract: Panoramic radiographs, also known as orthopantomograms, are commonly utilized by dentists to gain a comprehensive understanding of the patient’s oral health and perform orthodontic procedures. However, due to physician burnout and time constraints, many dentists may use them hastily which could result in medical negligence. To streamline the workflow for dentists, we establish a mission to segment five oral structures on panoramic radiographs, namely Alveolarcrest, Condyle, Neuraltube, Sinusmaxillaris and Teeth. A Cascaded Multi-scale Mask2former(CMMask2former) method is proposed for this task. For small objects, we design a multi-scale masked attention specifically for the mask area. The entire structure is designed in a two-stage cascade for localization and prediction. Our results demonstrate superior predictive performance compared to other methods.