A Two-Stage Multi-modal Fusion Framework for Robust Key-point Detection of Wisdom Teeth with Panoramic Radiographs

The study proposes a two-stage multi-modal fusion framework for key point detection of impacted wisdom teeth with panoramic radiographs(OPGs). It is crucial for the clinical diagnosis and surgical planning of impacted wisdom teeth to get accurate localization of anatomical key points in panoramic radiographs. However, this task continues to present a significant challenge to clinical automated systems. The existing end-to-end deep learning models often result in unreliable precision due to the large variations in tooth morphology and orientation, particularly in severely impacted cases. The proposed solution framework consists of two stages. In the initial stage, a segmentation network offers a robust region proposal, effectively eliminating background interference. Subsequently, the study introduces an innovative multi-modal input fusion strategy, where a lightweight pose estimation network is trained on a synthesized 3-channel input consisting of a grayscale image (texture), a binary segmentation mask (shape prior), and a zero-padding channel. This approach explicitly guides the model using precise shape information, encouraging it to learn rotation-invariant geometric features. The proposed method was validated on a clinical data-set with 200 patches of panoramic dental radiographs. The empirical research results demonstrate that the framework achieves state-of-the-art performance, reaching a mean Average Precision (mAP@.5:.95) of 0.918 for key point estimation. This significantly outperforms standard end-to-end baselines, demonstrating the effectiveness and practicality of the proposed multi-modal, segmentation-guided approach for high-precision medical image analysis.