Proj 08 – Contoral Health – Dry Mouth Device

Comfort in Every Moment

Introducing Proj 08 – Contoral Health – Dry Mouth Device

DentiSeg leverages AI to analyse 3D dental scans, giving dentists a clearer picture of a patient’s dentition. It identifies present and missing teeth, then uses those findings to recommend capsule placement areas, making it especially useful in the management of xerostomia.

Team members

Sharan Shekaran (ESD), Laverne Ang Qin Xuan (EPD), Chu Mei Qi (ISTD), Asher Yeo Qiheng (ESD), Raymond Khan (ISTD), Lim Chun Yang Samuel (EPD), Benjamin Lee Ze Perng (ISTD)

Instructors:

Qin Yanxia

Writing Instructors:

Belinda Seet

Overview

An AI-powered tool that segments and colour-codes 3D dental scans, identifies present and missing teeth using FDI numbering, and suggests potential capsule placement areas — displayed through an interactive .obj viewer.

Processing pipeline

The following describes how DentiSeg processes a 3D dental scan from upload to output.

Platform Overview

Main Model

Evaluation & Results

The platform allows users to upload and analyse 3D dental scans in .obj format. Once a scan is uploaded, a trained AI model predicts tooth labels, and the predictions are grouped into individual teeth. Each tooth is assigned an FDI tooth number, and its vertex count is used to classify it as present or missing. The results are displayed as a color-coded segmented scan in an interactive 3D viewer, alongside a tooth-by-tooth summary. From this, the platform identifies present and missing teeth and suggests possible capsule placement areas.

The selected model is DentiSeg Pro, a fine-tuned version of TSegNet [ToothGroupNetwork by Ho Yeon Lim and Min Chang Kim — https://github.com/limhoyeon/ToothGroupNetwork], adapted for improved performance on dental scans with multiple missing teeth. Among the three models evaluated — TSegNet (baseline), DentiSeg Plus (intermediate improved model), and DentiSeg Pro (final fine-tuned model), DentiSeg Pro achieved the best performance on the Contoral benchmark.

Rather than training from scratch, DentiSeg Pro was fine-tuned from the DentiSeg Plus checkpoint using a dataset of 72 jaw scans, all of which were partial-edentulism cases, with 71 of the 72 scans containing three or more missing teeth. This focused the fine-tuning specifically on challenging missing-tooth scenarios, which was one of the team’s main technical contributions.

It uses a two-stage design made up of Farthest Point Sampling (FPS) and Boundary-Detail Learning (BDL). FPS captures the overall structure of the full dental scan, while BDL refines local tooth boundaries in more difficult regions. This helps the model handle complex scans more effectively, particularly those involving multiple missing teeth.

After segmentation, the platform displays the FDI tooth number, vertex count, and presence status for each expected tooth. A threshold of 500 vertices is used to determine whether a tooth is classified as present or missing.

For evaluation, the team manually assessed each company-provided test scan to record which teeth were present or missing as ground-truth data. The model outputs were then compared against this ground-truth on a tooth-by-tooth basis. The evaluation focuses on present-or-missing tooth detection rather than exact segmentation boundary quality.

Current Limitations

The platform’s evaluation currently focuses on present-or-missing tooth detection and does not fully measure segmentation boundary quality, meaning a scan may score well numerically even if some boundaries appear visually imperfect. The capsule placement feature is a geometric planning aid and not a clinically validated treatment recommendation system. The platform is designed to assist dentists rather than replace professional judgement.

Project Vision:

DentiSeg aims to build a more intelligent dental analysis workflow grounded in strong tooth segmentation. The long-term goal is to improve model performance on difficult scans, particularly those involving partial tooth loss or irregular spacing, and to strengthen how segmentation results can support automated capsule placement suggestions. More broadly, DentiSeg demonstrates how AI can bridge the gap between raw model outputs and a clinically relevant, usable interface for dental professionals.