Electronic Theses and Dissertation

Abstrak - Karies gigi merupakan penyakit kronis yang memerlukan deteksi dini untuk mendukung penanganan yang tepat. Pemanfaatan deep learning pada citra klinis gigi intraoral menjadi salah satu pendekatan yang menjanjikan, namun variasi kualitas citra dan kondisi pencahayaan yang tidak merata sering memengaruhi kinerja model, sehingga tahap pra-segmentasi berupa filtering perlu diterapkan untuk menormalkan pencahayaan dan meningkatkan kualitas citra sebelum diproses lebih lanjut. Penelitian ini bertujuan untuk mengevaluasi pengaruh teknik filtering pra-segmentasi terhadap kinerja segmentasi dan klasifikasi karies gigi berbasis deep learning. Teknik filtering yang digunakan meliputi Discrete Cosine Transform (DCT), Wavelet Filtering, dan Homomorphic Filtering yang diterapkan menggunakan pendekatan photometric normalization. Segmentasi dilakukan menggunakan arsitektur U-Net dengan backbone ResNet-50 dan VGG19, sedangkan klasifikasi menggunakan ResNet-50. Hasil penelitian menunjukkan bahwa DCT menghasilkan nilai SSIM tertinggi (0,7827) yang mencerminkan preservasi struktur citra terbaik, sedangkan Wavelet Filtering menghasilkan nilai entropy tertinggi (7,8446). Pada tahap segmentasi, kombinasi VGG19-DCT menghasilkan performa terbaik dengan Test IoU 0,9479 dan Test Dice Score 0,9549. Pada tahap klasifikasi, seluruh model dengan filtering berhasil melampaui model tanpa filtering sebagai pembanding, dengan accuracy terbaik sebesar 80,21% dan F1-score terbaik sebesar 83,89% yang dicapai oleh model dengan segmentasi backbone VGG19-Homomorphic. Meskipun penerapan filtering pra-segmentasi tidak memberikan kontribusi yang signifikan terhadap kinerja segmentasi, penerapannya terbukti meningkatkan performa klasifikasi karies gigi secara konsisten.

Kata kunci: karies gigi, deep learning, filtering citra, discrete cosine transform, wavelet filtering, homomorphic filtering

Abstact – Dental caries is a chronic disease that requires early detection to support appropriate clinical management. The application of deep learning to intraoral clinical dental images has shown promising results; however, variations in image quality and uneven illumination conditions often affect model performance, making pre-segmentation filtering necessary to normalize illumination and improve image quality before further processing. This study aims to evaluate the impact of pre-segmentation filtering techniques on the performance of deep learning-based dental caries segmentation and classification. The filtering techniques applied include Discrete Cosine Transform (DCT), Wavelet Filtering, and Homomorphic Filtering, implemented using a photometric normalization approach. Segmentation was performed using a U-Net architecture with ResNet-50 and VGG19 backbones, while classification was conducted using ResNet-50. The results show that DCT achieved the highest SSIM value (0.7827), reflecting the best structural preservation, while Wavelet Filtering produced the highest entropy value (7.8446). In the segmentation stage, the VGG19-DCT combination achieved the best performance with a Test IoU of 0.9479 and a Test Dice Score of 0.9549. In the classification stage, all models with filtering outperformed the model without filtering used as a reference, with the best accuracy of 80.21% and the best F1-score of 83.89% achieved by the model with VGG19-Homomorphic segmentation backbone. Although the application of pre-segmentation filtering did not contribute significantly to segmentation performance, it consistently improved the classification performance of dental caries detection. Keywords: dental caries, deep learning, image filtering, discrete cosine transform, wavelet filtering, homomorphic filtering