Multi-view deep learning for classification of maxillary canine impaction severity using CBCT-derived images
This study developed and evaluated a multi-view DenseNet-121 deep learning model using CBCT-derived images from four perspectives to accurately classify maxillary canine impaction severity, achieving 91.2% overall accuracy and demonstrating its potential to support early orthodontic intervention.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the complex landscape of human dentition, the upper canine teeth hold a unique and often troublesome position. These teeth, which are essential for tearing food and shaping the smile, are the second most likely to get stuck after the wisdom teeth. Because they travel a long and winding path to reach their final spot, they frequently become trapped within the jawbone or gum tissue, a condition known as impaction. When this happens, the tooth cannot emerge on its own, and if left untreated, it can cause cysts to form, damage the roots of neighboring teeth, or disrupt the alignment of the entire dental arch. To solve this, orthodontists rely on imaging to see exactly where the tooth is and how stuck it is. While standard two-dimensional X-rays are common, a more detailed three-dimensional scan called cone-beam computed tomography, or CBCT, offers a clearer view of the tooth's position in space. However, interpreting these detailed scans requires significant time and expertise, and the radiation involved means doctors must be careful about when to use them.
A team of researchers set out to see if artificial intelligence could help navigate this diagnostic challenge. They developed a computer system designed to look at CBCT scans and automatically determine how severe a canine impaction is. Instead of trying to process the entire three-dimensional volume of the jaw at once, which is computationally heavy, the researchers took a different approach. They sliced the 3D scan into four familiar two-dimensional views that dentists already use: a panoramic view of the whole jaw, a frontal view from the front, a side view, and a cross-sectional slice through the tooth. They then trained a deep learning model, a type of computer program inspired by the human brain, to look at each of these four pictures separately and decide if the tooth was erupting normally, mildly stuck, or severely stuck.
The study focused on 74 patients, examining 148 upper canine teeth in total. For each tooth, the researchers generated 592 distinct images across the four different views. Human experts first labeled these images, categorizing each tooth as having normal eruption, mild impaction, or severe impaction based on specific angles and overlaps visible in the scans. The computer model was then taught to recognize these patterns. When the researchers tested the model, they found that looking at all four views together was far more effective than looking at just one. By combining the information from the panoramic, frontal, side, and cross-sectional images, the system correctly classified the severity of the impaction 91.2% of the time. This multi-view approach proved superior to any single image type; for instance, while the panoramic view alone was quite good at 88.5% accuracy, the frontal view alone struggled, achieving only 74.3% accuracy.
The results showed that the system was particularly reliable at identifying the two extremes: teeth that were coming in normally and teeth that were severely stuck. It correctly identified normal cases 93.4% of the time and severe cases 91.5% of the time. Crucially, the model never made the dangerous mistake of calling a severely stuck tooth "normal" or a normal tooth "severe." The most difficult category for the computer was the middle ground of mild impaction, where it was correct about 77% of the time. This is likely because mild cases look very similar to normal teeth on an X-ray, making them harder to distinguish even for humans. The researchers noted that the system paid attention to the exact same parts of the tooth that a human specialist would, focusing on the tip of the canine and the roots of the neighboring teeth, which suggests the computer was learning the right features rather than guessing.
This work demonstrates that artificial intelligence can serve as a powerful assistant in orthodontics, capable of synthesizing multiple angles of a scan to provide a consistent and rapid assessment of tooth impaction. While the model is not a replacement for a dentist, it offers a way to screen patients more efficiently and flag cases that need immediate attention. The researchers emphasized that for this tool to be used in real-world clinics, it will need to be tested on larger groups of people from different locations and with different types of scanners. Until then, the system remains a promising proof of concept, showing that by breaking a complex three-dimensional problem into familiar two-dimensional pieces, computers can learn to see dental issues with a high degree of accuracy, potentially helping to catch problems earlier and guide better treatment plans.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.