A dual-source intraoral image dataset for robust dental crack detection in complex real-world and home-care scenarios

Dental cracks represent a prevalent and elusive clinical challenge, often leading to severe tooth fractures if undetected. While artificial intelligence (AI) offers promising diagnostic support, the automated detection of micro-cracks in real-world environments remains a significant bottleneck. Existing public dental datasets predominantly focus on macroscopic diseases captured under strictly controlled clinical settings, restricting the development of robust algorithms for teledentistry. To bridge this gap, we introduce a dual-source, expert-annotated intraoral image dataset specifically designed for robust dental crack detection in complex scenarios. The dataset comprises 1,474 high-resolution images acquired via smartphones and home-use intraoral cameras under visible light. Each image features meticulous bounding-box annotations by experienced dental specialists using LabelMe. To facilitate advanced domain adaptation and robust model training, the dataset preserves complex real-world interferences (e.g., severe glare, saliva, and blurring) and provides standardized annotations in YOLO, PASCAL VOC, and COCO formats. This dataset establishes a rigorous benchmark for developing highly robust AI models for automated dental crack detection in everyday home-care environments.