DESIGN OF AN INTELLIGENT AUXILIARY GRADING SCHEME FOR DENTAL CARIES BASED ON DENTAL IMAGE SEGMENTATION AND FEATURE QUANTIFICATION

Abstract

Dental caries screening in primary care is challenged by small early-stage lesions, blurred boundaries, variable image quality, and subjective interpretation. This paper proposes an intelligent auxiliary grading scheme based on TransUNet and feature quantification. After grayscale normalization, contrast-limited adaptive histogram equalization, Gaussian denoising, and quality assessment, the model combines convolutional local-feature extraction with Transformer-based global context modeling to generate pixel-level lesion probability maps and binary masks. Connected-region analysis then quantifies lesion area, location, contour, overlap, and regional confidence. A shared classification branch produces C0-C3 auxiliary grades, while probability entropy, mask morphology, image-quality flags, and lesion confidence jointly trigger clinician review. The scheme defines patient-level dataset separation, annotation cross-checking, comparative and ablation experiments, and auditable outputs linking each prediction to its source image, preprocessing parameters, model version, mask, features, grade probabilities, and review status. Dice, IoU, precision, recall, accuracy, and F1-score are specified as validation metrics. The resulting framework is interpretable and traceable, but remains an auxiliary screening design requiring clinical calibration and external validation.