ViSTooth: A Visualization Framework for Tooth Segmentation on Panoramic Radiograph

TL;DR

ViSTooth addresses the challenge of accurate tooth segmentation on panoramic radiographs by combining a Mask R‑CNN–based automatic segmentation with a visualization-driven human–machine workflow. The framework introduces domain-specific metrics, glyph-based feature representations, and a low-dimensional distribution view to guide expert corrections and iterative retraining. Case studies and an expert study demonstrate improved IoU from $75.14\%$ toward $80.11\%$ with expert-driven feedback, while users report high usability and usefulness of the visual interface. The approach promises more reliable segmentation and serves as a foundation for integrating segmentation with downstream dental disease diagnosis and treatment planning.

Abstract

Tooth segmentation is a key step for computer aided diagnosis of dental diseases. Numerous machine learning models have been employed for tooth segmentation on dental panoramic radiograph. However, it is a difficult task to achieve accurate tooth segmentation due to complex tooth shapes, diverse tooth categories and incomplete sample set for machine learning. In this paper, we propose ViSTooth, a visualization framework for tooth segmentation on dental panoramic radiograph. First, we employ Mask R-CNN to conduct preliminary tooth segmentation, and a set of domain metrics are proposed to estimate the accuracy of the segmented teeth, including tooth shape, tooth position and tooth angle. Then, we represent the teeth with high-dimensional vectors and visualize their distribution in a low-dimensional space, in which experts can easily observe those teeth with specific metrics. Further, we expand the sample set with the expert-specified teeth and train the tooth segmentation model iteratively. Finally, we conduct case study and expert study to demonstrate the effectiveness and usability of our ViSTooth, in aiding experts to implement accurate tooth segmentation guided by expert knowledge.

Figures (6)

• Figure 1: The pipeline of VisTooth for tooth segmentation on panoramic radiograph.
• Figure 2: The illustration of Mask R-CNN.
• Figure 3: The extracted feature exploration component. The expert can start analysis from (A)the overview of the dataset. The (B)feature glyph, the (C)zoomed view and the (D)similarity view provide detailed information for feature exploration.
• Figure 4: (A) displays the process of expert correction. (B) captures instances where individual variations lead to scattered outliers.
• Figure 5: Cluster A exhibited a mixed distribution pattern. Cluster B consistently differentiated into five major distribution patterns.