tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Patch-Based detection of proximal caries on bitewing radiographs

Authors :

Sana Esmaeili¹ Parnian Alizadeh oskoee² Tahmineh Razi³ Asiyeh Dadghar⁴ Kasra Rahimipour⁵ Ata Jodeiri⁶

1- دانشگاه علوم پزشکی تبریز 2- دانشگاه علوم پزشکی تبریز 3- دانشگاه علوم پزشکی تبریز 4- دانشگاه علوم پزشکی تبریز 5- دانشگاه علوم پزشکی تبریز 6- دانشگاه علوم پزشکی تبریز

Keywords :

Proximal Caries،Bitewing Radiographs،Deep Learning،Patch-based classification،Medical image analysis

Abstract :

Abstract—Proximal caries lesions, also known as interproximal caries, were identified as cavities forming on the contact surfaces between adjacent teeth—areas challenging to clean and prone to early-stage decay. This study proposed a deep learning-based framework for automatic proximal caries detection using bitewing radiographs. A patch-based classification strategy was employed to localize lesion-centered regions that are often overlooked in global image analysis. To enhance diagnostic performance, an attention-based Multiple-Instance Learning (MIL) approach was applied to aggregate patch-level features into robust image-level predictions. We trained and evaluated the system on a dataset of 1,084 bitewing radiographs. The proposed MIL model achieved a test accuracy of 93.1%, significantly outperforming both global image classification (64.1%) and patch-only methods (66.2%). These results demonstrated the effectiveness of attention-based MIL in learning fine-grained features associated with caries. The system was designed to support diagnostic decision-making and facilitate early, non-invasive intervention in clinical dental practice.