A TWO-LEVEL METHOD FOR COMPLEX DIGITAL IMAGE RECOGNITION USING ROI MASKING AND A HYBRID U-NET 3+ / RESNET-U-NET ARCHITECTURE ON CASE STUDY OF DENTAL CARIES DETECTION ON X-RAY IMAGES

Abstract

The task of recognizing complex digital images is an important and relevant challenge in many scientific and technical domains. It has particular significance in medical diagnostics based on the analysis of radiographic digital images. The application of artificial intelligence systems provides a foundation for automating this process effectively. However, in several cases—particularly for the early detection of dental caries from radiographic dental images—standard digital image recognition methods fail to deliver accurate results. Increasing the accuracy of such diagnostics requires combining efficient segmentation algorithms for dental radiographs with adaptive deep learning models capable of analyzing only the relevant image regions.

This paper presents a two-level method for automatic recognition of complex digital images, demonstrated through the task of detecting carious lesions on dental X-ray images. The proposed system combines a segmentation stage with target analysis within a defined region of interest (ROI).

At the first stage, a modified U-Net 3+ architecture is applied, integrating a lightweight MobileNet encoder and Vision Mamba Layers in the decoder. This configuration ensures high segmentation accuracy and stability while reducing computational complexity. The resulting mask of the dental region is used to form the ROI, which constrains further analysis to clinically significant areas of the image while excluding background, soft tissues, and artifacts.

At the second stage, caries detection is performed using an optimized ResNet-U-Net model that combines residual blocks with a U-shaped structure, enabling precise reconstruction of object boundaries even under low-contrast conditions. Experimental studies conducted on the DENTEX dataset confirmed the effectiveness of the proposed approach: Dice = 0.74, Precision = 0.88, and Recall = 0.90, which outperform the baseline U-Net, U-Net 3+, and MobileMamba-U-Net models. The analysis demonstrated that ROI masking significantly reduces the number of false positives, improves model stability under varying exposure conditions, and decreases inference time by approximately one-third.

The proposed approach provides a balance between accuracy and computational efficiency, forming the basis for developing intelligent information systems capable of recognizing complex digital images. As shown by the conducted research, it is particularly suitable for intelligent decision-support systems in dental diagnostics. The method has potential for integration into software and mobile applications for automated screening of dental radiographs. Future research may focus on expanding the training dataset, implementing multi-pathology learning, and adapting the approach to other types of radiographic data.