Raith Stefan, Pankert Tobias, de Souza Nascimento Jônatas, Jaganathan Srikrishna, Peters Florian, Wien Mathias, Hölzle Frank, Modabber Ali
Department of Oral and Maxillofacial Surgery, RWTH Aachen University Hospital, Pauwelsstraße 30, 52074, Aachen, Germany.
Inzipio GmbH, Krantzstr. 7 Building 80, 52070, Aachen, Germany.
Sci Rep. 2025 Jan 7;15(1):1097. doi: 10.1038/s41598-024-83031-0.
For the planning of surgical procedures involving the bony reconstruction of the mandible, the autologous iliac crest graft, along with the fibula graft, has become established as a preferred donor region. While computer-assisted planning methods are increasingly gaining importance, the necessary preparation of geometric data based on CT imaging remains largely a manual process. The aim of this work was to develop and test a method for the automated segmentation of the iliac crest for subsequent reconstruction planning.
A total of 1,398 datasets with manual segmentations were obtained as ground truth, with a subset of 400 datasets used for training and validation of the Neural Networks and another subset of 177 datasets used solely for testing. A deep Convolutional Neural Network implemented in a 3D U-Net architecture using Tensorflow was employed to provide a pipeline for automatic segmentation. Transfer learning was applied for model training optimization. Evaluation metrics included the Dice Similarity Coefficient, Symmetrical Average Surface Distance, and a modified 95% Hausdorff Distance focusing on regions relevant for transplantation.
The automated segmentation achieved high accuracy, with qualitative and quantitative assessments demonstrating predictions closely aligned with ground truths. Quantitative evaluation of the correspondence yielded values for geometric agreement in the transplant-relevant area of 92% +/- 7% (Dice coefficient) and average surface deviations of 0.605 +/- 0.41 mm. In all cases, the bones were identified as contiguous objects in the correct spatial orientation. The geometries of the iliac crests were consistently and completely recognized on both sides without any gaps.
The method was successfully used to extract the individual geometries of the iliac crest from CT data. Thus, it has the potential to serve as an essential starting point in a digitized planning process and to provide data for subsequent surgical planning. The complete automation of this step allows for efficient and reliable preparation of anatomical data for reconstructive surgeries.
在涉及下颌骨骨重建的外科手术规划中，自体髂嵴移植与腓骨移植一样，已成为首选的供区。虽然计算机辅助规划方法越来越重要，但基于CT成像的几何数据的必要准备在很大程度上仍是一个手动过程。这项工作的目的是开发和测试一种用于髂嵴自动分割的方法，以便进行后续的重建规划。
总共获得了1398个带有手动分割的数据集作为基准真值，其中400个数据集的子集用于神经网络的训练和验证，另一个177个数据集的子集仅用于测试。采用在3D U-Net架构中使用TensorFlow实现的深度卷积神经网络来提供自动分割的管道。迁移学习用于模型训练优化。评估指标包括骰子相似系数、对称平均表面距离和专注于与移植相关区域的修正95%豪斯多夫距离。
自动分割实现了高精度，定性和定量评估表明预测结果与基准真值紧密对齐。对对应关系的定量评估得出，在移植相关区域的几何一致性值为92%±7%（骰子系数），平均表面偏差为0.605±0.41毫米。在所有情况下，骨骼都被识别为具有正确空间方向的连续物体。两侧髂嵴的几何形状都能持续且完整地被识别，没有任何间隙。
该方法成功地从CT数据中提取了髂嵴的个体几何形状。因此，它有可能成为数字化规划过程中的一个重要起点，并为后续的手术规划提供数据。这一步骤的完全自动化能够高效且可靠地为重建手术准备解剖数据。
