Efficient CBCT Segmentation via nnU-Net and Structure-Aware Post-processing

Changkai Ji, Yusheng Liu, Yuxian Jiang, Lisheng Wang

Keywords: nnU-Net, Structure Aware Post-processing, Computational efficiency

TL;DR: This paper presents an nnU-Net-based method with Structure Aware Post-processing and optimized inference strategies for efficient and accurate multi-class anatomical structure segmentation in CBCT.

Abstract: Accurate segmentation of anatomical structures from cone-beam computed tomography (CBCT) is essential for clinical applications in dentistry, maxillofacial surgery, and orthodontics. The ToothFairy3 Challenge has a comprehensive 77-class segmentation task, emphasizing both accuracy and computational efficiency. In this work, we present a method based on the nnU-Net framework, enhanced with a Structure Aware Post-processing (SAP) strategy. nnU-Net serves as a backbone for multi-class segmentation, while SAP refines predictions by introducing individualized thresholds for each anatomical structure, thereby mitigating noise and preserving clinically important fine structures. To further improve efficiency, we disabled mirroring augmentation during training and employed inference acceleration strategies, including the removal of test-time augmentation and optimized interpolation on floating-point tensors. Experimental results validate the effectiveness of our approach in balancing segmentation accuracy with computational efficiency.

Submission Number: 15