Electronic Theses and Dissertation

Karies gigi merupakan salah satu penyakit mulut paling umum dan dapat menyebabkan komplikasi serius jika tidak ditangani. deteksi dini sangat penting untuk memungkinkan intervensi yang efektif, menurunkan biaya perawatan, serta mencegah kerusakan lebih lanjut. kemajuan terbaru dalam bidang deep learning telah memungkinkan proses deteksi karies secara otomatis melalui citra klinis. namun, sebagian besar pendekatan yang ada masih mengandalkan citra dengan pemrosesan minimal, yang masih mengandung struktur tidak relevan dalam proses diagnosis, seperti lidah, bibir, dan gusi, yang dapat mempengaruhi akurasi diagnosis. penelitian ini mengembangkan model segmentasi gigi berbasis u-net untuk meningkatkan performa klasifikasi karies gigi menggunakan arsitektur resnet-50, resnext-50, dan inceptionv3. model segmentasi dilatih dengan transfer learning menggunakan backbone resnet-50, vgg19, dan inceptionv3, dievaluasi dengan metrik iou dan dice score. selanjutnya, model klasifikasi dilatih menggunakan citra asli dan tersegmentasi secara terpisah dengan hyperparameter yang sama. perbandingan performa dilakukan menggunakan confusion matrix dan analisis confidence interval. visualisasi grad-cam digunakan untuk mengevaluasi fokus model dalam proses pengambilan keputusan. hasil penelitian menunjukkan peningkatan performa yang konsisten pada semua model ketika segmentasi diterapkan, dengan akurasi tertinggi oleh resnext-50 sebesar 79,17%. visualisasi grad-cam mengonfirmasi bahwa segmentasi membantu model fokus pada area gigi yang relevan, meningkatkan akurasi dan keandalan klasifikasi. namun, analisis confidence interval menunjukkan bahwa meskipun terdapat peningkatan performa secara konsisten pada seluruh metrik, perbedaan yang diamati tidak signifikan secara statistik. kata kunci: karies gigi, klasifikasi karies, segmentasi gigi, deep learning, u-net.

Dental caries is one of the most common oral diseases and can lead to serious complications if left untreated. Early detection is essential to enable effective intervention, reduce treatment costs, and prevent further damage. Recent advancements in deep learning have made it possible to automatically detect caries through clinical images. However, most existing approaches still rely on minimally processed images, which often contain irrelevant structures for diagnosis, such as the tongue, lips, and gums, potentially affecting diagnostic accuracy. This study developed a U-Net-based tooth segmentation model to enhance caries classification performance using ResNet-50, ResNeXt-50, and InceptionV3 architectures. The segmentation models were trained using transfer learning with ResNet-50, VGG19, and InceptionV3 as backbones, and their performance was evaluated using IoU and Dice Score metrics. Subsequently, classification models were trained separately on original and segmented images with the same hyperparameters. Performance comparisons were conducted using confusion matrices and confidence interval analysis. Grad-CAM visualization was employed to evaluate the model's focus during the decision-making process. The results showed consistent performance improvements across all models when segmentation was applied, with ResNeXt-50 achieving the highest accuracy of 79.17%. Grad-CAM visualization confirmed that segmentation helped the models focus on relevant tooth areas, thereby improving classification accuracy and reliability. However, confidence interval analysis indicated that while performance improved consistently across all metrics, the observed differences were not statistically significant. Keywords: Dental Caries, Caries Classification, Tooth Segmentation, Deep Learning, U-Net.