Deteksi Kondisi Gigi Manusia Dari Citra Intraoral Menggunakan Model Ensemble WBF Pada YOLOv5 Dan YOLOv8

Syarif, Hisyam (2026) Deteksi Kondisi Gigi Manusia Dari Citra Intraoral Menggunakan Model Ensemble WBF Pada YOLOv5 Dan YOLOv8. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231035-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Kesehatan gigi sangat penting dalam kehidupan manusia, dan deteksi dini kerusakan gigi bisa menjadi tantangan tersendiri. Teknologi visi komputer telah digunakan untuk mendeteksi kondisi gigi menggunakan gambar intraoral, meningkatkan diagnosis dini dan perencanaan perawatan. Model deteksi objek berbasis deep learning seperti YOLOv5 dan YOLOv8 telah menunjukkan kinerja tinggi dalam mendeteksi objek gigi dalam gambar medis. Namun, efektivitas pendeteksian sangat dipengaruhi oleh teknik pasca-pemrosesan. Penelitian ini mengusulkan penerapan teknik ensemble post-processing Weighted Boxes Fusion (WBF) pada model YOLOv5 dan YOLOv8 untuk meningkatkan kinerja deteksi dan mengurangi deteksi false positive. Dataset yang digunakan berjumlah 2495 citra intraoral berasal dari Roboflow yang sudah memiliki anotasi dan memiliki 4 kelas yaitu caries, cavity, crack, dan tooth. Dataset tersebut akan digunakan untuk melatih dan menguji model, mengatasi masalah deteksi overlap pada gambar gigi yang kompleks. Model ensemble WBF diterapkan dengan menggabungkan prediksi bounding box dari kedua model YOLOv5 dan YOLOv8 untuk mengatasi masalah deteksi overlap. Setelah dilakukan pengujian model, dilakukan evaluasi metrik seperti precision, recall, F1-score dan mAP. Hasil evaluasi dari ensemble WBF akan dibandingkan dengan teknik post-processing lainnya, seperti Non-Maximum Suppression (NMS) dan Soft-NMS. Dari hasil penelitian, model yang diusulkan mencapai hasil uji terbaik dengan nilai mAP@0.5 sebesar 66,14%, nilai Precision sebesar 66,47%, dan IoU sebesar 90,83%. Sementara itu, model YOLOv5 dan YOLOv8 secara individual hanya memperoleh nilai mAP@0.5 dalam kisaran 36 37% dan IoU sebesar 39-41%. Model ensemble WBF yang diusulkan mampu memberikan peningkatan dan efektivitas dalam mendeteksi kondisi gigi pada citra intraoral.

==================================================================================================================================

Dental health is very important in human life, and early detection of dental damage can be a challenge in itself. Computer vision technology has been used to detect dental conditions using intraoral images, improving early diagnosis and treatment planning. Deep learning-based object detection models like YOLOv5 and YOLOv8 have shown high performance in detecting dental objects in medical images. However, the effectiveness of detection is greatly influenced by post processing techniques. This study proposes the application of the ensemble post-processing technique Weighted Boxes Fusion (WBF) on the YOLOv5 and YOLOv8 models to improve detection performance and reduce false positive detections. The dataset used consists of 2,495 intraoral images sourced from Roboflow, which are already annotated and have 4 classes: caries, cavity, crack, and tooth. The dataset will be used to train and test the model, addressing the issue of overlap detection in complex dental images. The WBF ensemble model is applied by combining the bounding box predictions from both the YOLOv5 and YOLOv8 models to address the overlap detection issue. After conducting model testing, metrics such as precision, recall, F1-score, and mAP were evaluated. The evaluation results of the WBF ensemble were compared with other post-processing techniques, such as Non-Maximum Suppression (NMS) and Soft-NMS. From the research results, the proposed model achieved the best test results with a mAP@0.5 value of 66.14%, a Precision value of 66.47%, and IoU of 90.83%. Meanwhile, the individual YOLOv5 and YOLOv8 models only obtained mAP@0.5 values in the range of 36-37% and IoU values of 39-41%. The proposed WBF ensemble model was able to provide improvements and effectiveness in detecting dental conditions in intraoral images.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Deteksi Objek, Gigi, Intraoral, WBF, YOLOv5, YOLOv8 Dental, Intraoral, Object Detection, WBF, YOLOv5, YOLOv8
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Hisyam Syarif
Date Deposited:	28 Jan 2026 02:49
Last Modified:	28 Jan 2026 02:49
URI:	http://repository.its.ac.id/id/eprint/130325

Actions (login required)

View Item