Proximal dental caries detection using CNNs and level set based crown extraction

Abstract

Proximal dental caries are dental cavities occurring on occluded surfaces between adjacent teeth. Proximal dental caries are diagnosed with dental x-ray images since they can hardly be observed with naked eye. However, a poor quality of dental x-ray images caused by uneven exposure, x-ray scattering, and a large variation makes it difficult for dentists to diagnose proximal dental caries. Therefore, we propose an automatic proximal dental caries detection system for intraoral (IO) and panoramic x-ray images. The system consists of probability generation, crown extraction, and refinement. In probability generation, a probability map of proximal dental caries is produced by pixel-wise convolutional neural networks (CNNs). In crown extraction, crown areas of teeth where dental caries possibly exist are extracted. In order to extract each crown separately, modified seams are used to isolate each tooth and a gum line is detected using the difference of mean intensity values between local areas above and below the gum line. Then, an edge-based level set method with regularization is utilized to extract crown areas. In refinement, the probability map is refined using crown areas to improve detection results. Crown areas and distance probabilities modeled by crown contours are used to eliminate false positives. Finally, the refined probability map is binarized to decide proximal caries regions. Experiments on both IO and panoramic images reveal that the proposed system using both CNNs and crown extraction is superior to $na\ddot{i}ve$ CNNs. Another contribution of this paper is that it is the first time to propose an algorithm for detection of proximal dental caries to the best of our knowledge.