Related Experiment Video
Updated: Jul 12, 2026
Midface Hypoplasia and Cranial Base Morphology in Syndromic Craniosynostosis: A Comparative Analysis Study Using a Predictive Regression Model
Published on: November 4, 2025
A Foundation Model for Generalisable Detection of Maxillary Sinus Abnormalities: A Multicentre and Clinical
Shijie Chen1, Rihui Song2, Peisheng Zeng1
1Guangdong Provincial Key Laboratory of Stomatology and Guangdong Provincial Clinical Research Center of Oral Diseases, Hospital of Stomatology, Guanghua School of Stomatology, Sun Yat-Sen University, Guangzhou, China.
Objectives:
Maxillary sinus abnormalities reduce quality of life and can be life-threatening. However, traditional supervised deep learning models for their detection are constrained by class-specific training, reliance on large-scale labelled datasets, and unpredictable generalisability. This study aims to achieve intelligent detection of complex and variable maxillary sinus abnormalities via self-supervised learning.
Method:
A maxillary sinus foundation model (MSFound) was developed using self-supervised learning on 30 794 unlabeled maxillary sinus images. Clinical validation of reconstructed images was conducted to verify the pre-training effect. MSFound was then fine-tuned using labelled images at different proportions for downstream tasks, including mucosal thickening, polypoid lesions, and the palatonasal recess. Model performance was evaluated on multicentre datasets using AUROC, AUPR, and accuracy, and was compared with different training strategies. Clinical applicability was also prospectively evaluated in multicentre settings.
Results:
MSFound successfully reconstructed the main anatomical structures of the maxillary sinus, and clinical validation confirmed the effectiveness of pre-training. During the fine-tuning process, MSFound achieved the highest performance with the least amount of labelled data compared to control groups across all tasks. The model consistently achieved the highest AUROC, AUPR, accuracy, and F1-score on multicentre test sets. Visualisation analyses showed that MSFound focused on key image regions when predicting correctly. Prospective clinical applicability results further demonstrated the real-world effectiveness of MSFound.
Conclusion:
MSFound enables robust detection of multiple maxillary sinus abnormalities with minimal labelled data, demonstrating strong generalisability and real-world clinical value. This study provides a self-supervised learning framework and a real-world evaluation paradigm for clinical dentistry AI.
