A Foundation Model for Generalisable Detection of Maxillary Sinus Abnormalities: A Multicentre and Clinical...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oral and Maxillofacial Radiology

Background:

• Maxillary sinus abnormalities significantly impact patient quality of life and can pose serious health risks.
• Traditional supervised deep learning methods for detecting these abnormalities face limitations due to class-specific training, extensive labeled data requirements, and variable generalizability.
• Self-supervised learning offers a promising alternative to overcome these challenges in medical image analysis.

Purpose of the Study:

• To develop and validate an intelligent system for detecting complex and diverse maxillary sinus abnormalities using self-supervised learning.
• To assess the model's performance and generalizability across multicenter datasets.
• To evaluate the clinical applicability of the developed model in real-world settings.

Main Methods:

• A maxillary sinus foundation model (MSFound) was created using self-supervised learning on over 30,000 unlabeled maxillary sinus images.
• The model underwent clinical validation through image reconstruction to confirm pre-training efficacy.
• MSFound was fine-tuned with varying proportions of labeled data for tasks like detecting mucosal thickening, polypoid lesions, and palatonasal recess, with performance evaluated using AUROC, AUPR, and accuracy.

Main Results:

• MSFound effectively reconstructed key maxillary sinus anatomical structures, validating the pre-training phase.
• The model achieved superior performance with significantly less labeled data compared to control groups across all detection tasks.
• Consistent high scores in AUROC, AUPR, accuracy, and F1-score were observed on multicenter test sets, indicating robust detection and strong generalizability.

Conclusions:

• The MSFound model demonstrates robust and generalizable detection of various maxillary sinus abnormalities using minimal labeled data, highlighting its clinical value.
• This self-supervised learning framework provides a novel approach for AI in clinical dentistry.
• The study establishes a practical evaluation paradigm for AI in real-world dental applications.