Arian Md Sahadul Hasan, Sifat Faisal Ahmed, Ahmed Saif, Mohammed Nabeel, Farook Taseef Hasan
Department of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh.
Adelaide Dental School, University of Adelaide, Adelaide, SA5000, Australia; Research and Innovations, Dental Loop Pty Ltd, Adelaide, SA5000, Australia.
Int J Med Inform. 2025 Mar;195:105769. doi: 10.1016/j.ijmedinf.2024.105769. Epub 2024 Dec 19.
The automated segmentation of individual teeth from 3D models of the human dental arch is challenging due to variations in tooth alignment, arch form and overall maxillofacial anatomy. Domain adaptation is a specialised technique in deep learning which allows models to adapt to data from different domains, such as varying tooth and dental arch forms, without requiring human annotations.
This study aimed to segment individual teeth from various dental arch morphologies in 3D intraoral scans using domain adaptation.
Twenty scanned dental arches from various age groups and developmental stages were used to generate 20 simplified synthetic variants of the scans. These synthetic variants, along with 16 natural scanned dental arches, were used to train the deep learning models. Domain adaptation was employed using Gradient Reversal Layer and Siamese Network techniques. The PointNet and PointNet++ model backbones were trained to align the latent space distribution of real and synthetic domains. Validations were performed on four unseen natural scanned arches, with and without domain adaptation enabled, to evaluate whether a 3D deep neural network can be trained without any human-annotated 3D models.
PointNet and PointNet++ models demonstrated a mean intersection-over-union between 0.34 and 0.36 mIoU without domain adaptation enabled and 0.80 and 0.95 mIoU, respectively with domain adaptation enabled when assessing natural scanned dental arches.
Domain adaptation techniques can enable training a segmentation deep learning model using synthetically generated 3D jaw scans without requiring human operators annotating the training data.
由于牙齿排列、牙弓形态和整体颌面解剖结构的差异，从人类牙弓的三维模型中自动分割出单个牙齿具有挑战性。域适应是深度学习中的一种专门技术，它允许模型适应来自不同域的数据，例如不同的牙齿和牙弓形态，而无需人工注释。
本研究旨在使用域适应技术在三维口腔内扫描中分割出各种牙弓形态的单个牙齿。
使用来自不同年龄组和发育阶段的20个扫描牙弓生成20个简化的合成扫描变体。这些合成变体与16个自然扫描牙弓一起用于训练深度学习模型。使用梯度反转层和连体网络技术进行域适应。训练PointNet和PointNet++模型主干以对齐真实域和合成域的潜在空间分布。在启用和未启用域适应的情况下，对四个未见的自然扫描牙弓进行验证，以评估是否可以在没有任何人工注释的三维模型的情况下训练三维深度神经网络。
在评估自然扫描牙弓时，PointNet和PointNet++模型在未启用域适应时的平均交并比在0.34至0.36 mIoU之间，启用域适应时分别为0.80至0.95 mIoU。
域适应技术可以在无需人工操作员注释训练数据的情况下，使用合成生成的三维颌骨扫描来训练分割深度学习模型。
