Deep learning-based segmentation of the maxillary sinus on panoramic radiographs using MedSAM and DeepLabv3+

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Panoramic radiographs are used in dental practice because of low radiation dose, patient comfort, and rapid acquisition. However, segmentation of the maxillary sinus remains difficult because of superimposed anatomical structures. We evaluated the performances of MedSAM and DeepLabv3+, two advanced segmentation models, for delineating the maxillary sinus on panoramic radiographs. Methods A total of 1,046 panoramic radiographs were retrospectively collected from a dental hospital and a private clinic. Maxillary sinus boundaries were manually annotated by two oral and maxillofacial radiologists and one general dentist, using the VGG Image Annotator. The dataset was randomly divided into training, validation, and test sets in a 60:20:20 ratio. Binary masks were generated, and segmentation was performed using MedSAM and DeepLabv3 + in Python. DeepLabv3 + was trained from scratch, while MedSAM was fine-tuned from pretrained weights. Model performance was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), precision, recall, F1-score, Normalized Surface Distance (NSD), and 95th percentile Hausdorff Distance (HD95). Results Both MedSAM and DeepLabv3 + demonstrated high segmentation performance for maxillary sinus delineation on panoramic radiographs. MedSAM achieved significantly higher overall segmentation performance than DeepLabv3+, with superior DSC (0.9570 ± 0.0169 vs. 0.9534 ± 0.0101, p < 0.001), IoU (0.9183 ± 0.0299 vs. 0.9124 ± 0.0185, p < 0.001), recall ( p < 0.001), F1-score ( p < 0.001), and HD95 (0.0161 ± 0.0081 vs. 0.0199 ± 0.0053, p < 0.001). DeepLabv3 + demonstrated significantly higher precision ( p = 0.002). Although MedSAM provided superior segmentation accuracy, DeepLabv3 + required substantially fewer model parameters and floating-point operations, indicating greater computational efficiency. Conclusions Although MedSAM demonstrated superior segmentation performance, this advantage should be interpreted in light of its substantially higher computational cost. DeepLabv3 + achieved high segmentation performance with considerably greater computational efficiency, highlighting the importance of balancing segmentation accuracy and computational practicality when selecting models for clinical application. These findings demonstrate the technical feasibility of deep learning–based maxillary sinus segmentation on panoramic radiographs. Nevertheless, further validation using diverse institutions, imaging protocols, and pathological conditions is required before routine clinical implementation.