| dc.description.abstract | Image segmentation in dentistry, both in clinical practice and research, is increasingly vital. With the rapid advancement of computer vision, deep learning applications proliferate in medical image analysis, enhancing decision-making in clinical environments. However, the adoption of AI in dentistry remains limited due to data and methodological constraints. This thesis presents significant advances in dental image segmentation, focusing on deep learning to equip dentists with efficient tools for accurate tooth region identification, covering fundamental data aspects and sophisticated methodologies. For 2D panoramic dental images, a segmentation approach based on a Generative Adversarial Network is introduced, called ToothPix. This method surpasses traditional segmentation techniques like U-Net, effectively capturing complex dental structures and extracting detailed tooth boundaries from even imprecise annotations. The lightweight discriminator within the GAN imposes realistic constraints on the generator, steering it towards authentic data representations. Experimental results illustrate that the ToothPix method demonstrates superior performance in the 2D tooth segmentation, outperforming existing methods with its highest Dice coefficient of 94.86% and an impressive Intersection over Union score of 89.06% . Its innovative approach yields marked improvements in both precision and sensitivity, as evidenced by its peak scores. This approach offers novel generative perspectives for tooth region segmentation. Addressing data scarcity and extending the 2D method to 3D, a comprehensive annotated dental CT dataset with corresponding benchmarks is proposed. Initially, a diverse 3D CBCT dental dataset from clinical cases is compiled, called CTooth. Subsequently, three segmentation benchmarks for the proposed dental dataset are established: a fully supervised model utilizing extensive labeled data, an active-learning model that iteratively engages specialists in the selection and annotation of unlabeled data, and a semi-supervised model that integrates both labeled and unlabeled data. The study also explores attention-focused methodologies in 3D tooth segmentation, expanding the understanding of various segmentation strategies. For the fully supervised benchmark, experimental results illustrate that the proposed BM-Unet shows superiority compared with other attention mechanisms and 3D UNet variants on the precision and distance metrics. For the active learning and semi-supervised benchmarks, the study extends state-of-the-art methodologies, applies them to the CTooth dental dataset, and conducts a comparative performance analysis, culminating in the demonstration of three-dimensional tooth modeling outcomes. Additionally, a clinic-friendly semi-supervised 3D dental segmentation method is introduced, called ACT-Tooth, based on insights from the 3D dataset and the semi-supervised benchmark. This approach innovatively combines 3D convolutional neural networks and 3D transformer mechanisms, a first in the semi-supervised 3D dental segmentation field. Experimental results demonstrate the ACT-Tooth method achieves the highest Dice Similarity Coefficient of 89.13% and an exceptional Intersection over Union of 79.12% , exceeding some recent methods. The ACT-Tooth addresses the challenge of limited annotated data, reducing reliance on extensive manual annotations and maintaining high segmentation accuracy. In conclusion, this research makes significant contributions to dental image segmentation, providing innovative approaches to address the prevalent challenges of data insufficiency and methodological constraints in dental AI applications. It pioneers the use of Generative Adversarial Networks for 2D dental imaging and innovatively extends some insights to 3D applications, utilizing a comprehensive dataset and benchmarks. This advancement notably enhances diagnostic and therapeutic capabilities in dentistry. Moreover, the integration of supervised, active-learning, and semi-supervised models demonstrates adaptability across a spectrum of clinical settings, underscoring the transformative potential of AI in augmenting the accuracy and efficiency of dental care. These developments collectively underscore AI's integral role in advancing modern dentistry, particularly in refining treatment precision and operational efficiency. | en_US |
