Hegyi Alexandra, Somodi Kristóf, Pintér Csaba, Molnár Bálint, Windisch Péter, García-Mato David, Diaz-Pinto Andres, Palkovics Dániel
1 Semmelweis Egyetem, Fogorvostudományi Kar, Parodontológiai Klinika Budapest, Szentkirályi u. 47., 4. em., 1088 Magyarország.
2 Empresa de Base Tecnológica Internacional de Canarias, S. L. (EBATINCA) Las Palmas de Gran Canaria Spanyolország.
Orv Hetil. 2024 Aug 11;165(32):1242-1251. doi: 10.1556/650.2024.33098.
Introduction: The goal of segmentation is to reconstruct cone-beam computed tomography (CBCT) images in three dimensions (3D). In oral surgery and periodontology, digital data processing enables 3D planning of surgical interventions. Commonly used threshold-based segmentation is fast but inaccurate, whereas semi-automatic methods are sufficiently accurate but time-consuming. Recently, with artificial intelligence-based technologies, automatic segmentation of CBCT images has become feasible. Objective: To present a deep learning segmentation model trained on CBCT images derived from clinical practice and to evaluate its efficiency. Method: The study consisted of three phases: establishing the training dataset, training the deep learning model and testing its accuracy. CBCT images of 70, partially edentulous patients were used to establish the training dataset. The deep learning model, based on the SegResNet architecture, was developed within the MONAI framework. To verify the accuracy of the deep learning model, 15 CBCT scans were used processed using the deep learning-based segmentation and semi-automatic segmentation, and the results were compared. Results: The similarity between the two methods, based on intersection over union, was on average 0.91 ± 0.02. The average Dice similarity coefficient was 0.95 ± 0.01, and the average Hausdorff (95%) distance was 0.67 mm ± 0.22 mm. There was no statistically significant difference in the volume of the 3D models segmented by the deep learning architecture compared to those created by semi-automatic segmentation (p = 0.31). Discussion: The deep learning model used in our study performed segmentation of CBCT images with accuracy comparable to other artificial intelligence-based systems reported in the literature. Since the CBCT images were sourced from routine clinical practice, the deep learning model segmented periodontal bone topography and alveolar ridge defects with relatively high reliability. Conclusion: The deep learning model accurately segmented the mandible in dental CBCT scans. Therefore, the deep learning-based 3D models could be suitable for digital planning of reconstructive oral and periodontal surgical interventions. Orv Hetil. 2024; 165(32): 1242–1251.
分割的目标是在三维（3D）空间中重建锥形束计算机断层扫描（CBCT）图像。在口腔外科和牙周病学中，数字数据处理能够实现手术干预的三维规划。常用的基于阈值的分割方法速度快但不准确，而半自动方法虽然足够准确但耗时。最近，随着基于人工智能的技术发展，CBCT图像的自动分割已变得可行。目的：展示一个在源自临床实践的CBCT图像上训练的深度学习分割模型，并评估其效率。方法：该研究包括三个阶段：建立训练数据集、训练深度学习模型并测试其准确性。使用70例部分牙列缺失患者的CBCT图像建立训练数据集。基于SegResNet架构的深度学习模型在MONAI框架内开发。为验证深度学习模型的准确性，使用15例CBCT扫描，分别采用基于深度学习的分割和半自动分割进行处理，并比较结果。结果：基于交并比的两种方法之间的相似度平均为0.91±0.02。平均Dice相似系数为0.95±0.01，平均豪斯多夫（95%）距离为0.67mm±0.22mm。与半自动分割创建的三维模型相比，深度学习架构分割的三维模型体积无统计学显著差异（p = 0.31）。讨论：我们研究中使用的深度学习模型对CBCT图像进行分割的准确性与文献中报道的其他基于人工智能的系统相当。由于CBCT图像源自常规临床实践，深度学习模型对牙周骨形态和牙槽嵴缺损的分割具有较高的可靠性。结论：深度学习模型在牙科CBCT扫描中准确分割了下颌骨。因此，基于深度学习的三维模型可能适用于口腔和牙周重建手术干预的数字规划。《匈牙利医学周报》。2024年；165（32）：1242–1251。
