Predicting Orthognathic Surgery Necessity in Class III Malocclusion Using Deep Learning

This thesis explores the use of deep learning, specifically Convolutional Neural Networks (CNNs), to enhance the diagnosis and treatment planning for Class III malocclusion. This condition, affecting approximately 5\% of the global population, is marked by a skeletal discrepancy between the upper and lower jaws, often requiring orthognathic surgery or orthodontic interventions. Current clinical approaches often depend on subjective interpretations of radiographic and photographic images, making accurate diagnosis challenging. CNNs offer the potential to improve diagnostic precision by automatically identifying patterns in complex image data, providing a more objective method. The study used a dataset of radiographs, profile photographs, and intraoral images of Class III malocclusion patients to train CNN models, particularly based on the ResNet architecture. These models were optimized through data augmentation and hyperparameter tuning to increase accuracy. Results demonstrated that CNNs significantly exceeded expectations, with high accuracy in distinguishing between patients requiring surgery and those manageable with orthodontic treatment alone. Additionally, CNNs reduced the time needed for image analysis and decreased the subjectivity of manual diagnosis. ResNet models, known for their ability to handle deep and complex learning tasks, were particularly effective. A key aspect of this research was comparing CNNs with traditional machine learning techniques. While traditional models like Support Vector Machines (SVM) and Random Forests performed well, they required manual feature extraction, which CNNs avoided by learning directly from images. This automated approach streamlined the diagnostic process. The findings underscore the potential of CNNs to transform orthodontic care by providing more accurate, consistent, and efficient diagnostic tools for Class III malocclusion. The study also acknowledges some limitations, particularly regarding the size and quality of available datasets, which may affect the generalizability of the results. In conclusion, this thesis highlights CNNs as a promising tool for diagnosing Class III malocclusion, offering improvements in diagnostic accuracy and efficiency. The integration of CNNs into clinical practice could enhance treatment outcomes and support more reliable, data-driven decision-making in orthodontics.