Beyond the Naked Eye: A Unified Deep Learning Approach for Precision Dental Diagnosis

At Nik Fanavar Nasir Co., we believe that the future of medicine lies in the seamless integration of intelligent systems with clinical routine. Today, we are excited to share the details of our latest research paper, "Advanced Deep Learning-Based Approach for Tooth Detection, and Dental Cavity and Restoration Segmentation in X-Ray Images", developed by our technical team at Smartory Labs.

This research represents the algorithmic core of our Smarteeth platform, addressing a critical challenge in dentistry: consistency. Since dentists vary in their ability to identify caries based on experience and fatigue, we developed a unified method to standardize and enhance diagnostic accuracy.

The Challenge: Complexity in Radiography

Dental X-rays are dense with information. A major issue in previous AI studies was the inability to effectively handle different types of X-rays (like Panoramic vs. Bitewing) or accurately number teeth due to the visual symmetry of the jaw. Furthermore, class imbalance—where healthy teeth far outnumber cavities in training data—often skewed results.

The Solution: A Novel Unified Pipeline

Our paper introduces a two-stage pipeline that mimics how a meticulous expert analyzes an image. Instead of looking at the whole mouth at once for cavities, our method breaks it down:

1. Intelligent Detection & Numbering (The "Where")

First, we utilize a customized YOLOv8 model to detect individual teeth. However, AI often confuses the left and right sides of the mouth due to symmetry. To solve this, we introduced a novel post-processing technique. This algorithm automatically corrects predicting errors by analyzing the position of teeth, achieving a detection precision of over 90%. It accurately assigns a Universal Teeth Numbering (UTN) label to every tooth.

2. Precise Segmentation (The "What")

Once a tooth is detected, the system crops it out. By focusing on individual teeth rather than the entire X-ray, our segmentation model can find minute details. This allows for the precise outlining (segmentation) of:

• Cavities (Caries)
• Restorations (Fillings)

This targeted approach led to a 6.8% improvement in overall accuracy (mAP@50) compared to state-of-the-art models.

Clinical Relevance & Utility

This study focuses on addressing practical diagnostic challenges by proposing a versatile and efficient pipeline.

Multi-View Support:

The method is engineered to process various radiographic types—Panoramic, Bitewing, and Periapical—within a single framework.

Operational Efficiency:

By automating routine tasks such as tooth numbering and initial screening, the system aims to optimize time and resource consumption in clinical workflows.

Diagnostic Consistency:

The model acts as an objective assistant, helping to reduce inter-examiner variability and standardize the assessment process.

Scientific Validation & Application

Publishing these findings serves to transparently validate the technical algorithms powering the Smarteeth platform. By utilizing recognized standard datasets (such as Tufts) and rigorous evaluation methods, we aim to bridge the gap between experimental models and clinical requirements. This approach ensures that the solutions developed at Nik Fanavar Nasir Co. are built on a solid technical foundation, capable of serving as reliable and precise tools in the medical diagnostic process.

Read the full technical paper presented at the 11th RSI International Conference on Robotics and Mechatronics (ICRoM 2023): IEEE Xplore