Fully Automated Mandibular Condyle Segmentation: More Detailed Extraction With Hybrid Customized SAM

Accurate segmentation of the mandibular condyle is a key step in three-dimensional reconstruction, which is clinically crucial for digital surgical planning in oral and maxillofacial surgery. Quantitative analysis of its volume and morphology can provide an objective basis for preoperative assessment and postoperative efficacy evaluation. Although many deep learning-based approaches have achieved remarkable success, several challenges persist. Current methods are constrained by low-resolution global image maps, produce masks with blurred boundaries, and require large datasets to ensure accuracy and robustness. To address these challenges, we propose a novel framework for condylar segmentation by adapting the “Segmentation Anything Model” (SAM) to cone beam computed tomography (CBCT) imaging data, with targeted architectural optimizations to enhance segmentation accuracy and boundary delineation. Our framework introduces two novel architectural components: (1) a dual-adapter system combining feature augmentation and transformer-level prompt enhancement to improve target-specific contextual learning, and (2) a boundary-optimized loss function that prioritizes anatomical edge fidelity. For clinical practicality, we further develop ConDetector to enable fully automated prompting without manual intervention. Through extensive experiments, we have shown that our adapted SAM (using Ground Truth as a prompt) achieves state-of-the-art performance, reaching a Dice coefficient of 94.73% on a relatively small sample set. The fully automated SAM even achieves the second-best segmentation performance, with a Dice coefficient of 94.00%. Our approach exhibits robust segmentation capabilities and achieves excellent performance even with limited training data.