MSLPNet: multi-scale location perception network for dental panoramic X-ray image segmentation

Tooth segmentation, as one of the key techniques of medical image segmentation, can be widely applied to various medical applications, e.g., orthodontic treatment, corpse identification, dental training systems, dental disease diagnosis, etc. Although there have been many studies on tooth segmentation, few tooth segmentation studies have focused on enhancing tooth segmentation with fuzzy root boundaries which is a difficult but essential task in dentistry to determine the root resorption and the tooth brace. The existing methods for tooth segmentation usually exploited the contrast enhancement to sharpen the boundaries of teeth, while the final segmentation results heavily depended on the subsequent processing steps of the method. To address the issue of fuzzy boundaries, this paper proposes a novel multi-scale location perception network to segment teeth from panoramic X-ray images. The core of the proposed method stems from three aspects: (1) multi-scale structural similarity loss conducts accurate prediction with clear boundary from patch-level scale; (2) location perception module locates each tooth pixel in the image from the perspective of global-level scale; (3) aggregation module reduces the semantic gap between multi-scale feature branches. Our proposed method was tested on a dataset containing 1500 dental panoramic X-ray images and the Dice score of our method is 93.01%, which outperforms the state-of-the-art approaches. Besides, our proposed method has better boundary quality and the Pratt (1978)’s figure of merit used for boundary quality evaluation reaches 76.56%.