Access to advanced imaging technologies has revolutionized the field of dentistry, particularly concerning the maxillary sinus anatomy, which is closely related to dental procedures. Recent research from the University Dental Hospital, Sharjah, UAE, has spotlighted the accessory ostium (AO)—an anatomical variant of the maxillary sinus—through the lens of artificial intelligence (AI) and deep learning models.
The study aimed to assess the accuracy of deep learning algorithms, particularly the ResNet-101V2 model, in detecting the presence of accessory ostia using coronal cone beam computed tomographic (CBCT) images. A comprehensive analysis involved collecting 454 coronal images from 856 large field of view (FOV) CBCT scans extracted from the hospital's dental radiology archives between January 1 and June 30, 2024.
The research is grounded on the clinical significance of AOs, which may vary widely and are often linked with sinus pathology. Accessory ostia can lead to complications such as sinusitis and the so-called two-hole syndrome—where mucous drainage becomes obstructed—resulting from the altered sinus ventilation and drainage pathways. Understanding and detecting these anatomical variations is pivotal for dental professionals, especially when considering implants or surgical interventions.
To bridge the gap between anatomical knowledge and clinical practice, the researchers utilized pre-trained deep learning models—specifically Visual Geometry Group (VGG16), MobileNetV2, and ResNet-101V2—as the foundation for their classification efforts. After evaluating these models, the team determined ResNet-101V2 provided the most promise for accuracy.
The researchers employed various data augmentation techniques to increase their dataset's size from 454 images to 1260, thereby enhancing the model’s learning capacity and performance. On implementing L1 (Lasso regression) regularization, the researchers aimed to mitigate the risk of overfitting—an important consideration when working with machine learning algorithms.
The findings anchored the study on solid metrics. The ResNet-101V2 model achieved test accuracy and loss rates of 81% and 51%, respectively. Additional performance indicators included precision, recall, F1-score, and area under the curve (AUC) values of 0.82, 0.81, 0.81, and 0.87 respectively. The robustness of the model's performance was reflected not only through these statistics but also through inter-rater reliability scores, which showed substantial agreement at 0.87 between the two examiners involved.
This level of inter-rater variance affirms the reliability of AI-assisted diagnostics. The intra-rater reliability, which measures consistency over time by the same assessors, was commendably high—ranging from 0.91 to 0.95. These findings are compelling, considering the common challenges associated with radiographic interpretation, where subjective assessments can often lead to discrepancies.
The study highlights the paradigm shift occurring within radiology and dentistry through AI applications—where tools can drastically reduce reading time and improve accuracy. The practical application of these technologies suggests not only enhanced experiences for patients but also significant advancements for practitioners who rely on precise imaging to inform their treatment plans.
Despite the promising results, the researchers acknowledged limitations tied to sample size and the scope of images analyzed. Future investigations are needed to explore the efficacy of these models on three-dimensional CBCT scans rather than the two-dimensional slices utilized in this study. Three-dimensional imaging could potentially capture more complex anatomical relationships and improve diagnostic capabilities even more.
Conclusively, the team's work establishes ResNet-101V2 as a formidable tool for detecting accessory ostia within coronal CBCT images. This serves as foundational research for subsequent AI initiatives aimed at refining imaging assessments and interpretations across dental radiology. Academic and clinical institutions may now advance toward creating comprehensive, accurate AI models to revolutionize patient care standards.