Bitewing radiographs play an important role in dental diagnosis, especially in the detection of caries, evaluation of restorations, and follow-up of endodontic procedures. However, their manual interpretation is time-consuming and subject to variability, particularly in cases involving overlapping structures, low contrast, and restorative materials. In this context, deep learning methods have emerged as promising tools for assisting radiographic analysis. This paper presents an applied semantic segmentation pipeline for bitewing radiographs based on U-Net with a ResNet50 backbone. The proposed approach combines grayscale preprocessing, data augmentation, transfer learning, and multi-class segmentation to identify crowns, restorations, and root canal treatments. Experimental evaluation yielded an Intersection over Union (IoU) of 0.43, precision of 0.47, recall of 0.98, and F1-score of 0.44. These results indicate high sensitivity but limited precision, mainly due to false positives. Therefore, the present study should be interpreted as an initial feasibility investigation rather than as a comparative benchmark or a clinically deployable solution. Even so, the model was able to identify relevant dental structures and provides an applied basis for discussing bitewing radiographs as Smart Medical Media, in which AI-generated overlays may enrich visual interpretation, documentation, and communication in digital dentistry.