Reproduce “Automatic tracing of mandibular canal pathways using deep learning” on OpenResearch

About this paper

Mrinal Kanti Dhar, Zeyun Yu

Researchers at the Big Data Analytics and Visualization Laboratory, University of Wisconsin-Milwaukee, developed an end-to-end 3D deep learning framework for automated mandibular canal tracing from CBCT data. The framework, which employs a novel centerline-based ground truth generation and a multi-stage post-processing pipeline, achieved an F1-score of 0.56 and an mIoU of 0.70, with a Mean Curve Distance from Prediction to Ground truth (MCD-PG) of 0.65 mm, demonstrating enhanced precision and reduced false positives.

Key insights

• Employing centerline-based ground truths significantly improves model precision, resulting in predictions with fewer false positives and better alignment to the actual canal anatomy compared to fixed-diameter tubular ground truths.
• A robust post-processing stage is essential for translating raw deep learning outputs into clinically usable results by effectively mitigating common artifacts like thickened predictions, false positives, and broken segments.
• Data augmentation via on-the-fly random rotations and specific 3D U-Net architectural adaptations (e.g., 2x2x1 max-pooling) are crucial for enhancing model generalization and robustness in volumetric medical image segmentation.

Summary by alphaXiv.