Diagnosis of Periodontal Disease in Dental Panoramic Radiographs using Few-Shot Learning CVAEs

다국어 초록 (Multilingual Abstract)

In the rapidly evolving field of medical imaging, deep learning systems have catalyzed remarkable advancements in various diagnostic tasks. However, these systems' reliance on sizable training datasets places them at a fundamental disadvantage. To address this critical bottleneck, researchers are actively exploring the paradigm of few-shot learning, which promises to mitigate the challenge of data scarcity by leveraging maximal information from minimal data samples. Few-shot learning algorithms are particularly pertinent in medical imaging scenarios, where labeling data requires high costs and open data sources are scarce, thereby limiting the volume of available training data. The effectiveness of few-shot learning in such contexts has the potential to be a transformative breakthrough. By articulating the challenges posed by few-shot classification in the face of limited data, this study puts forward a novel Few-shot learning algorithm, specifically tailored for the diagnosis and classification of prevalent periodontal diseases using a constrained dataset of dental panoramic radiographs. By employing the UNet architecture, regions of interest (ROI) were generated from a limited number of panoramic radiographs, which were subsequently processed using Convolutional Variational Autoencoder to extract latent features. These latent vectors were then subjected to a clustering algorithm to perform unsupervised clustering. Following this, a select set of labeled images was utilized to assign labels to images indicative of periodontal diseases, streamlining the diagnostic process despite the data limitations. To validate the efficacy of this framework, comparative analyses were conducted against the traditional supervised learning method with image augmentation and autoencoder-based latent vector clustering. The findings reveal that our approach notably surpasses conventional models in performance, emphasizing enhanced diagnostic precision and efficiency, even in data-constrained environments.

목차 (Table of Contents)

• List of Figures iii
• List of Tables iv
• Abstract v
• 1. Introduction . 1
• 2. Data Description 4
• 2.1. The Tufts Dental Database . 4
• 2.2. Hanyang University Seoul Hospital Database . 5
• 2.3. Data Augmentation . 6
• 3. framework . 8
• 3.1. Convolutional Neural Network 9
• 3.2. UNet 14
• 3.3. Convolutional Variational Autoencoder 15
• 3.4. Unsupervised Clustering 16
• 3.4.1. K-means Clustering . 17
• 3.4.2. DBSCAN . 17
• 3.4.3. Gaussian Mixture Model 17
• 3.5. Bayesian Optimization 17
• 4. Results and Discussion 20
• 4.1. Experimental Setup . 20
• 4.2. Comparative Performance Analysis 22
• 4.2.1. Tuffs Dental Database 22
• 4.2.2. Hanyang University Seoul Hospital Dental Database 24
• 5. Conclusion and Future Work 27
• References 29
• 국문요지 32