How erdanc/tooth-and-alveolar-bone-segmentation-from-cbct works
Project Type and Purpose
• This is an open-source AI pipeline (no UI) for fully automatic tooth and alveolar bone segmentation from cone-beam CT (CBCT) volumes.
• It comprises a multi-stage deep learning workflow: ROI (region of interest) localization, tooth detection, single-tooth segmentation, and instance-level post-processing.
File-Structure Analysis
• Top-level scripts:
• cnt_skl_dect.py (skeleton/count-skeleton detection)
• roi_dect.py (entry point for ROI localization)
• ins_tooth_seg.py (entry point for instance-level segmentation)
• test.py (inference orchestration)
• networks/
• Reusable VNet-based architectures for segmentation.
• roi_localization/
• dataloaders/ (binary mask generation, ToothLoader)
• networks/ (VNet model definition)
• train.py (ROI-localization training loop)
• utils/ (loss functions, learning-rate ramps, general utilities)
• tooth_detection/
• dataloaders/ (skeleton extraction, data generation, ToothLoader)
• networks/ (centroid_cluster, Simaese landmark net, VNet)
• train.py (tooth detection training loop)
• utils/ (losses, ramps, thin-plate spline, util)
• single tooth segmentation/
• dataloaders/ (singleToothLoader)
• networks/ (per-tooth segmentation architectures)
• train_singleTooth_segmentation.py
• utils/ (losses, ramps, thin-plate spline, util)
README Insights
• Deep-learning stack: PyTorch 1.4.0.
• Training and inference procedures split into stages, with separate scripts.
• Input: .nii.gz CBCT files; output: segmented masks for bone and individual teeth.