Imagine an emergency department physician working in the middle of the night, facing a stack of X-ray images that must be reviewed one by one. Tired eyes and declining concentration create an opportunity for errors. According to data cited in a recently published research protocol in BMJ Open, up to 3.1% of fractures are missed during a patient's initial visit to the emergency department. The figure may sound small, but the implications are not: delayed fracture diagnosis can worsen pain, prolong suffering, and reduce the long-term effectiveness of treatment.
The question, then, is whether machines can help.
A Vulnerable Region with Abundant Data
The dentomaxillofacial region—including the teeth, upper and lower jaws, cheekbones, orbits, and midface—is one of the areas of the body frequently affected by mechanical trauma. In 2019 alone, an estimated 178 million new fracture cases occurred worldwide, 10.7 million of which involved facial bones. The leading causes include road traffic accidents, physical violence, and falls.
Injuries in this region involve more than broken bones. Complications can include double vision (diplopia), partial facial nerve paralysis, impaired sense of smell, and the formation of an abnormal connection between the maxillary sinus and oral cavity—a condition known as an oroantral communication, which may progress to a fistula or chronic sinusitis if left untreated.
To diagnose these conditions, clinicians rely on radiographic imaging, ranging from affordable panoramic and periapical radiographs to highly accurate but more expensive CT scans that involve greater radiation exposure. CBCT has emerged as an increasingly popular middle ground. Yet all these modalities ultimately depend on one imperfect element: human interpretation.
A Machine Trained to See
drg. Silviana Farrah Diba, Sp.RKG, Subsp. RP (K), a researcher from the Department of Dental Radiology at the Faculty of Dentistry, Universitas Gadjah Mada, who is also pursuing a doctoral degree at the Faculty of Medicine, Public Health, and Nursing at UGM, led the development of this scoping review protocol together with a multidisciplinary team. The team includes experts in anatomy, medical radiology, electrical engineering and information technology, as well as health behaviour specialists—a collaboration reflecting the complexity of the problem being addressed.
The protocol published in BMJ Open is not a report of results but rather a roadmap: a systematic framework for mapping the extent to which global research has explored the use of artificial intelligence (AI) to detect dentomaxillofacial fractures on diagnostic images.
“Although previous reviews have examined the use of AI in various diagnostic imaging techniques, fractures in the dentomaxillofacial region have not previously been the specific focus of a review,” the researchers wrote in their paper.
The AI in question is not simply an ordinary computer program. In medical imaging, the most relevant technology is deep learning, a branch of machine learning that uses layered artificial neural networks to recognise patterns in complex data. One of the most widely used architectures is the convolutional neural network (CNN), which is specifically designed for image analysis. Models such as AlexNet, U-Net, DenseNet, and ResNet have already been tested for detecting mandibular fractures on panoramic radiographs and CT scans.
The process works by training AI using thousands of radiographic images labelled by experts. The system learns to recognise patterns—fracture lines, changes in bone density, and morphological abnormalities—and automatically identifies them in new images it has never encountered before.
A Roadmap Towards a Comprehensive Review
The protocol uses the Arksey and O'Malley framework, modified according to recommendations by Levac et al., and follows the PRISMA-ScR reporting guidelines. The literature search will cover seven major databases, including PubMed, Scopus, ScienceDirect, Cochrane Library, SpringerLink, IEEE, and ProQuest, covering publications from January 2000 to June 2023.
Two independent reviewers will screen titles and abstracts before proceeding to full-text assessment. In the event of disagreement, a third reviewer will be involved. A minimum agreement threshold of 75% was established before formal screening begins. The extracted data will cover three major categories: study characteristics; comparator characteristics, including whether AI was compared with specialist clinicians; and AI model characteristics, including the architecture used, who labelled the training data, and performance parameters such as sensitivity, specificity, and F1-score.
Interestingly, the protocol explicitly acknowledges its own limitations. The search is restricted to English-language articles, meaning studies from non-Anglophone countries—including Indonesia—may not be captured. Article quality will also not be assessed, as this is not a requirement for scoping reviews.
When Radiologists and Engineers Sit at the Same Table
What makes this protocol particularly noteworthy is not only the findings it may eventually produce, but also its approach. The research brings dental radiologists together with information systems engineers—two fields that rarely intersect directly in the same research setting.
Previous studies have shown that the number of studies investigating AI diagnostic accuracy has increased sharply since 2017. However, most have focused on fractures in other parts of the body or on dental conditions that do not specifically involve trauma. Dentomaxillofacial fractures—with all their anatomical complexity—remain an underexplored area in the global literature.
The planned scoping review is designed to fill that gap. Its findings are expected to provide a foundation for more specific research: perhaps clinical trials, perhaps the development of AI models tailored to patients in Southeast Asia, or perhaps clinical guidelines that could help the emergency physician working late at night—whose eyes are growing heavy—not miss even a single fracture line.
Source DOI: https://doi.org/10.1136/bmjopen-2022-071324
Authors: Anny Anggraini; drg. Achmad Zam Zam Aghasy, M.Kes.
Photo: Pexels