A High-Fidelity Dataset for Tooth Crack Detection in Complex Clinical Scenarios (DFD-Net)

Tooth cracks are considered the silent killer of natural dentition, characterized by irreversible structural destruction and high diagnostic invisibility. Early detection is paramount to saving teeth; however, identifying cracks is severely hindered by their sub-pixel width, low contrast, and high morphological similarity to physiological fissures. Crucially, the absence of standardized public benchmarks and the limitations of conventional isotropic convolutions in capturing directional linear features have rendered this field a no man’s land for automated diagnostics. To bridge this gap, we curated a high-fidelity clinical dataset comprising 753 images annotated by prosthodontists and proposed the Direction-Aware and Frequency-Aligned Network (DFD-Net).