Superpixel-Based Classification of Occlusal Caries Photography

Segmentation using superpixels of simple photographic image for classification of occlusal caries according to the lesion severity, makes significant changes to the way experts annotate the image, but also in the way of learning and evaluating of an automatic classifier. Working on an extension of the lower part of the 6-class ICDAS (International Caries Detection and Assessment System) scale, we are building a classifier exhibiting very low Random Forests OOB (Out-Of-Bag) Error estimation, without performing any image enhancement or morphological operation techniques. We also demonstrate the robustness of the classifier's performance over the size of superpixels by introducing a shrinking factor in the model's learning phase. Finally we highlight the complications to evaluate the models performance through the cross-validation procedure, arising from the class inequalities as distributed across the limited image dataset.