Impacted teeth, particularly third molars that fail to erupt properly due to insufficient space or improper angulation, represent a common dental condition that can lead to severe complications including infection, cyst formation, and damage to adjacent structures. Traditional diagnosis relies heavily on manual interpretation of panoramic dental X-ray images by clinicians, a process that is time-consuming, subject to human variability, and lacks pixel-level precision. This paper presents an AI-based impacted tooth detection system using the U-Net deep learning architecture, a convolutional neural network specifically designed for biomedical image segmentation. The proposed system performs pixel-level segmentation of impacted tooth regions from panoramic dental X-ray images, providing precise boundary delineation that conventional object detection methods cannot achieve. The system integrates data annotation, model training using PyTorch, and deployment via a Flask-based web application into a unified end-to-end pipeline. Preprocessing steps including grayscale conversion, resizing to 256 × 256 pixels, and pixel normalization ensure consistent input quality. The trained model achieved an overall segmentation accuracy of approximately 87%, with precision of 85%, recall of 89%, and an F1-score of 87%. Experimental results and confusion matrix analysis confirm that the proposed system reliably detects impacted tooth regions while maintaining a low rate of false predictions. The system demonstrates strong real-time performance through a user-friendly web interface, making it a practical diagnostic support tool for dental professionals.