Proximal Dental Caries Detection


Joonhyang Choi, Hyunjun Eun, and Changick Kim


본 연구에서는 


Proximal dental caries are diagnosed using dental X-ray images. Unfortunately, the diagnosis of proximal dental caries is often stifled due to the poor quality of dental X-ray images. Therefore, we propose an automatic detection system to detect proximal dental caries in periapical images for the first time. The system comprises four modules: horizontal alignment of pictured teeth, probability map generation, crown extraction, and refinement. We first align the pictured teeth horizontally as a pre-process to minimize performance degradation due to rotation. Next, a fully convolutional network are used to produce a caries probability map while crown regions are extracted based on optimization schemes and an edge-based level set method. In the refinement module, the caries probability map is refined by the distance probability modeled by crown regions since caries are located near tooth surfaces. Also we adopt non-maximum suppression to improve the detection performance. Experiments on various periapical images reveal that the proposed system using a convolutional neural network (CNN) and crown extraction is superior to the system using a na¨ıve CNN. 


  • Overall structure


  • The goal of the proposed system is to obtain a set of proximal dental caries regions C = {G1, G2, ..., GN }, where each group Gn denotes a single carious region. To achieve this goal, the system consists of four modules: HAPT, probability map generation, crown extraction, and refinement.



  • Horizontal alignment of pictured teeth (HAPT)

  • (a) Tooth top points on teeth areas (marked in greed). (b) Fitted top line (in red). (c) HAPT image.

    As a pre-processing stage of the proposed system, we horizontally align the pictured teeth and remove unnecessary regions. Although teeth are captured roughly upward in periapical images, we further align the pictured teeth horizontally to minimize performance degradation due to rotation. First, teeth areas are roughly extracted by Otsu’s thresholding. Then the set of tooth top points T = {t1, t2, ..., tw} can be determined by automatically selecting the most upper pixel of teeth areas in each column as shown in Fig. 3a. w is defined as the image width. In the Figure, the tooth top line ltop is modeled by the RANdom SAmple Consensus (RANSAC) based on T. The RANSAC randomly chooses samples for modeling and the model that represents all data best is selected after several trials. ltop is effectively modeled by the RANSAC because random sampling excludes outliers such as points near the gum line, while the least square method uses all data for modeling. Finally, the image is aligned by rotating the image by the angle of ltop. In addition, we crop the areas above the tooth top line and one third of the image height from the bottom since the areas usually do not contain crowns. The aligned and cropped image is represented in the Figure and we will call this image the HAPT image in this paper.


  • Probability map generation

  • (a) The architecture of our CNN. (b) Caries probability map generated from the CNN.

    In our proposed system, proximal dental caries regions are to be detected by binarization of a caries probability map. To this end, dense inference is carried out efficiently by a FCN to produce a caries probability map. Probability map generation comprises CNN training and caries probability estimation. In caries probability estimation, the FCN is utilized, which perform a fully convolutional process over the entire image instead of using patches in sliding windows to reduce inference time.


  • Crown extraction

  • (a) Teeth isolation. (b) Gum line detection. (c) Crown region segmentation.


    In crown extraction, we obtain crown regions by using optimization schemes and an edge-based level set method from the HAPT image. Crown extraction consists of teeth isolation, gum line detection, and crown region segmentation. In teeth isolation, we find lines that separate adjacent teeth. Then, the line that splits crowns and roots is detected in gum line detection. In crown region segmentation, crowns regions are segmented based on the isolation lines and the gum line with an edge-based level set method.



  • Refinement

  • Probability adjustment. (a) Test image. (b) Caries probability map from CNN. (c) Caries probability map by prior 1). (d) Caries probability map by prior 1) and 2).

    Non-maximum suppression. (a) A caries region (in green) and the local window (in blue). (b) Suppression result.

    In the refinement module, we boost the detection performance based on probability adjustment and non-maximum suppression. In probability adjustment, high probabilities in non-caries pixels are suppressed based on crown regions. In non-maximum suppression, we detect accurate caries regions among the binarized regions from the caries probability map.



  • Experimental results

  • Visual comparison. (a) Ground truth regions. (b) Detected proximal dental caries regions. (c) Caries probability maps.


    Performance evaluation. (a) Precision-recall curve. (b) F1 score against threshold.


    Joonhyang Choi, Hyunjun Eun, and Changick Kim, "Boosting Proximal Dental Caries Detection via Combination of Variational Methods and Convolutional Neural Network," Journal of Signal Processing Systems, vol. 90, no. 1, pp.87-97, Jan. 2018.