Electronic Theses and Dissertation

Karies merupakan salah satu masalah kesehatan gigi dan mulut yang masih banyak dijumpai dan memerlukan deteksi dini yang akurat untuk mencegah kerusakan jaringan gigi yang lebih parah. Pemanfaatan deep learning pada citra intraoral berpotensi membantu proses deteksi dan klasifikasi karies secara otomatis, namun kinerjanya masih terbatas akibat kurangnya kejelasan informasi batas dan kontur gigi, terutama pada tahap awal karies yang memiliki perbedaan visual yang tidak mencolok dengan gigi sehat. Penelitian ini bertujuan untuk menganalisis dampak penerapan edge detection pasca segmentasi terhadap kinerja klasifikasi karies gigi berbasis deep learning. Proses diawali dengan segmentasi area gigi menggunakan Feature Pyramid Network (FPN) dengan backbone EfficientNetB1, kemudian hasil segmentasi diproses menggunakan metode edge detection berbasis deep learning, yaitu Holistically-Nested Edge Detection (HED), Richer Convolutional Features (RCF), Dense Extreme Inception Network for Edge detection (DexiNed), dan Pixel Difference Network (PiDiNet), di mana setiap metode diuji secara terpisah dan dibandingkan pengaruhnya terhadap kinerja klasifikasi. Citra gigi asli, citra hasil segmentasi, serta citra hasil edge detection selanjutnya digunakan sebagai masukan bagi model klasifikasi berbasis DenseNet169, MobileNetV2, dan EfficientNetB1. Kinerja sistem dievaluasi menggunakan Intersection over Union (IoU) dan Dice Coefficient pada tahap segmentasi serta Confusion Matrix, Accuracy, Precision, Recall, Specificity, dan F1-score pada tahap klasifikasi. Hasil evaluasi menunjukkan bahwa penerapan
edge detection pasca segmentasi secara konsisten meningkatkan kinerja klasifikasi pada ketiga arsitektur yang diuji. Kombinasi terbaik dicapai oleh DenseNet169 dengan metode RCF, menghasilkan accuracy 89,06%, precision 88,73%, recall 82,89%, Specificity 93,10% dan F1-score 85,70%, membuktikan bahwa edge detection berbasis deep learning efektif meningkatkan akurasi sistem klasifikasi karies gigi berbasis deep learning.
Kata kunci: karies gigi, edge detection, HED, RCF, DexiNed, PiDiNet, klasifikasi citra.

Dental caries is one of the most prevalent oral health problems and requires accurate early detection to prevent further damage to dental tissue. The use of deep learning on intraoral images has the potential to assist in the automatic detection and classification of caries; however, its performance remains limited due to the lack of clarity in tooth boundary and contour information, particularly in the early stages of caries where visual differences from healthy teeth are not prominent. This study aims to analyze the impact of applying edge detection post-segmentation on the performance of deep learning-based dental caries classification. The process begins with tooth area segmentation using Feature Pyramid Network (FPN) with EfficientNetB1 as the backbone, followed by processing the segmentation results using deep learning-based edge detection methods, namely Holistically-Nested Edge Detection (HED), Richer Convolutional Features (RCF), Dense Extreme Inception Network for Edge Detection (DexiNed), and Pixel Difference Network (PiDiNet), where each method is tested separately and compared in terms of its effect on classification performance. The original tooth images, segmented images, and edge-detected images are then used as inputs for classification models based on DenseNet169, MobileNetV2, and EfficientNetB1. System performance is evaluated using Intersection over Union (IoU) and Dice Coefficient at the segmentation stage, as well as Confusion Matrix, Accuracy, Precision, Recall, Specificity, and F1-score at the classification stage. The evaluation results show that the application of edge detection post-segmentation consistently improves classification performance across all three architectures tested. The best combination is achieved by DenseNet169 with the RCF method, yielding an accuracy of 89,06%, precision of 88,73%, recall of 82,89%, Specificity of 93,10%, and F1-score of 85.70%, demonstrating that deep learning-based edge detection effectively enhances the accuracy of deep learning-based dental caries classification systems. Keywords: dental caries, edge detection, HED, RCF, DexiNed, PiDiNet, image classification.