Fuzzy K-Nearest Neighbor Based Dental Fluorosis Classification Using Multi-Prototype Unsupervised Possibilistic Fuzzy Clustering via Cuckoo Search Algorithm

Abstract

Dental fluorosis in children is a prevalent disease in many regions of the world. One of its root causes is excessive exposure to high concentrations of fluoride in contaminated drinking water during tooth formation. Typically, the disease causes undesirable chalky white or even dark brown stains on the tooth enamel. To help dentists screen the severity of fluorosis, this paper proposes an automatic image-based dental fluorosis segmentation and classification system. Six features from red, green, and blue (RGB) and hue, saturation, and intensity (HIS) color spaces are clustered using unsupervised possibilistic fuzzy clustering (UPFC) into five categories: white, yellow, opaque, brown, and background. The fuzzy k-nearest neighbor method is used for feature classification, and the number of clusters is optimized using the cuckoo search algorithm. The resulting multi-prototypes are further utilized to create a binary mask of teeth and used to segment the tooth region into three groups: white-yellow, opaque, and brown pixels. Finally, a fluorosis classification rule is created based on the proportions of opaque and brown pixels to classify fluorosis into four classes: Normal, Stage 1, Stage 2, and Stage 3. The experimental results on 128 blind test images showed that the average pixel accuracy of the segmented binary tooth mask was 92.24% over the four fluorosis classes, and the average pixel accuracy of segmented teeth into white-yellow, opaque, and brown pixels was 79.46%. The proposed method correctly classified four classes of fluorosis in 86 images from a total of 128 blind test images. When compared with a previous work, this result also indicates 10 out of 15 correct classifications on the blind test images, which is equivalent to a 13.33% improvement over the previous work.

Keywords: Dean’s index; Lévy flights; c-means clustering; cuckoo search; dental fluorosis; possibilistic.

Figure 1

Examples of each fluorosis class from the training set and their segmented images: (a) Image N2 (Normal); (b) Image F1_2 (Stage 1); (c) Image F2_2 (Stage 2); (d) Image F3_1 (Stage 3). The left column presents the expert’s labels of opaque pixels and brown pixels, encircled in black and red, respectively. The center column presents the predicted binary tooth masks. The right column presents the predicted white–yellow, opaque, and brown pixels in blue, green, and red colors, respectively.

Figure 2

Examples with correct prediction of each fluorosis class from the blind test set and their segmented images: (a) Normal; (b) Stage 1; (c) Stage 2; (d) Stage 3. The left column presents the expert’s labels of opaque pixels and brown pixels, encircled in black and red, respectively. The center column presents the predicted binary tooth masks. The right column presents the predicted white–yellow, opaque, and brown pixels in blue, green, and red colors, respectively.

Figure 3

Examples of misclassification on the blind test set and their segmented images. (a) Expert’s opinion: Normal; Prediction: Stage 3 with $r_{opaque}=19.84\%$ and $r_{brown}=3.23\%$. (b) Expert’s opinion: Stage 1; Prediction: Stage 2 with $r_{opaque}=33.95\%$ and $r_{brown}=0\%$.

Figure 4

An image in the class Normal was misclassified as Stage 3 if Line 3 in Algorithm 4 was removed, because the segmented image has $r_{opaque}=1.63\%$ (green color) and $r_{brown}=1.35\%$ (brown color).