A Novel Technique to Detect Caries Lesion Using Isophote Concepts

Abstract Background and Objectives Dental caries is one of the most common painful and infectious oral diseases. Early detection of caries lesion prevents the spreading of infection. Generally, dentists use x-ray images to locate the lesion's position. Dental x-ray images have poor intensity which results difficulties in finding exact affected area at a glance. Due to scarcity of dentists at government hospital, it becomes very difficult for the dentists to treat large number of patients in a short span of time. The objective of this work is to design a system to assist dentists to detect caries lesion quickly and more accurately. Deep learning based method is not suitable for this application because there is not enough training set is available to prepare the pre-trained model properly for deep learning. Traditional handcrafted method is desirable for such situation. Methods Dental x-ray not only contains the image of teeth and bonny structure of the jaws but also the tissues within the gum regions. So normal texture based segmentation is not enough to detect the caries lesion. In x-ray, caries lesion looks like a catchment basin, in which the depth at the center is maximum. Isophote along with geodesic active contour method is suitable to model such property of caries lesion. But prior to that multistage background elimination is essential to locate the suspected caries region. Ramdomness calculation and rescaling of that value on the basis of a small training data set is the first part of this multi stage background elimination process. Initial background elimination is performed on the basis of modified k-means clustering upon the entropy value and gray scale values of the x-ray image. In this clustering technique the number of cluster is determined automatically based on analyzing the distribution of data points. The clustering technique is immune against over clustering. Most of the caries lesion lies within the teeth region. Hence this region is surrounded by teeth region. This property is also checked to detect the suspected caries lesion and eliminate the background. Result Till now very limited dental x-ray databases with caries lesion is available online. ‘Digital dental periapical x-ray database for caries screening’ dataset is used to test the method. The proposed method achieved overall 94% of accuracy and average computational time is below 4.5 sec. Discussion This is an alternate solution to detect ROI when deep learning technique fails due to lack of exhaustive training set. This approach fails to generate correct result if resolution of the x-ray image is very low. Low resolution images make confusion between randomness and noise. In addition to that catchment basin properties are not identified properly. Due to this carries lesion are not properly identified.