Following the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) main theme, the Solvd research team proposes a novel approach to using Artificial Intelligence in analyzing 3D scans in dentistry. The goal is to precisely map the structure, position and roots of teeth while reducing noise in the image.
The global market for AI solutions in dentistry is estimated to reach $3,117.6 million by the end of 2034, compared with $421 million in 2024. The growth arises from, among other sources, using algorithms to aid in diagnostics, analysis, prevention and treatment planning.
The first step in nearly every diagnostic process in dentistry is identifying the dental structures in the jaw. The dentist needs not only to examine the teeth, but also to describe them in documentation, including the identification of diseases, malformations and patient-specific patterns.
Why applying AI in dentistry is challenging
Because every human is unique, spotting patterns and extracting them becomes especially challenging, as the line between the norm and abnormality can be thin. This is caused by several factors:
- Dense environment. The jaw contains a large number of interconnected elements, including teeth, gums, bone tissue, nerves and a plethora of other elements that may have an impact on the procedure. A scan may be completely incomprehensible to an untrained eye and traditional machine vision techniques get confused easily.
- Multiple elements impact the position of teeth. Previous procedures like tooth extraction, orthodontic issues and a lack of teeth due to injuries or bad health influence the jaw image and make it even more challenging to analyze.
- Natural diversity. People’s teeth are unique, so the machine learning tools may find it challenging to harvest information, as the norms in dentistry are broad and sometimes loosely defined. Also, the roots of teeth come in various unorthodox shapes, even for overall healthy teeth. This makes procedures like extraction more challenging, especially without the whole position and shape of the tooth fully mapped.
Yet applying AI in dentistry brings forth multiple benefits, including reduced costs and better patient recovery and experience. All healthcare requires good insight into the patient’s body (literally) and condition. As such, AI assistance in diagnostics or procedure preparation is a field where improving the algorithms and their performance may bring immediate benefits — the algorithms themselves are not making health decisions but rather provide information to humans to support better decision-making.
Our work
In pursuit of the benefits mentioned above, the Solvd research team, together with researchers from other facilities, delivered GEPAR3D (GEometric Prior-Assisted LeaRning for 3D). This novel approach unites the instance (tooth or other entity within jaw) detection with a segmentation, where particular elements and objects in the 3D image can be detected and classified.
The model maps a mouth’s 3d scan, where it uses the statistical position of teeth to find a central point to begin a search for teeth and map their presence or lack thereof, where applicable. Having the dental structure mapped, the system then classifies spotted entities, for example, recognizing implants, fractures or deformities.
Using position detection ensures the keeping of the anatomical context and consistency in the process, so the algorithm will not be misled by uncommon features or image flaws. Also, it saves time and computing power by giving the system a guiding light on the most probable position of a particular tooth. Nevertheless, the healthcare professional still needs to oversee the outcomes.
More details about the algorithm, the approach, and its internal workings can be found in the research paper published on ArXiv.
The effect
The system produces a 3D model of the teeth, with all features and shape clearly visible. This reduces the noise and background information from the scan, giving healthcare professionals a clear view of the teeth. This also includes, for example, the root structure, the potential entanglement of them, their position among other teeth or any other important information.
GEPAR3D shows a better performance than State-Of-The-Art methods of teeth positioning and segmentation. The method can find application in supporting better orthodontic planning and root resorption assessment.
The next step may be training the algorithm to work with children’s teeth, as the current version works with adult teeth only.
The research was delivered by a team consisting of Tomasz Szczepański, Szymon Płotka, Michal K. Grzeszczyk, Arleta Adamowicz, Piotr Fudalej, Przemysław Korzeniowski, Tomasz Trzciński and Arkadiusz Sitek, representing Solvd; the Sano Centre for Computational Medicine, Cracow, Poland; Jagiellonian University, Cracow, Poland; Jagiellonian University Medical College, Cracow, Poland; the Warsaw University of Technology, Warsaw, Poland; Research Institute IDEAS, Warsaw, Poland; and Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA.
The research will be presented at the upcoming 28th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2025), which will be held from Tuesday, September 23rd, to Saturday, September 27th, 2025, in Daejeon, Republic of Korea. More details about the conference can be found on the event website.
