Automated high-fidelity 3D reconstruction of middle-ear ossicles from low-resolution clinical CT using a deep...

This study validates an automated deep learning framework for generating high-fidelity 3D models of the middle-ear ossicles from low-resolution clinical CT images. The framework employs a sequential three-stage pipeline: (1) accurate Region of Interest (ROI) detection using YOLOv5x, (2) 4x super-resolution of the ROI with a Deep Back-Projection Network (DBPN), and (3) slice interpolation using a 2.5D U-Net to create a dense volumetric dataset. To ensure robust reconstruction from incomplete data, the interpolation stage integrates a "hint channel" that leverages anatomical priors. The framework demonstrated high accuracy, achieving a mean Average Precision (mAP50) of 0.9835 for ROI detection and producing final 3D models with a high degree of anatomical fidelity (Dice coefficient: 0.85; and mean surface distance: 4.8 µm). The hint channel's efficacy was most evident on an external inference set, where it successfully generated complete ossicular structures that were otherwise omitted due to sparse source information, demonstrating the model's strong generalization. Furthermore, the entire automated process, from CT scan to final 3D model, was completed within 5 min, offering a substantial improvement in workflow efficiency compared to manual methods that require approximately more than 20 min. The proposed framework is thus validated as a rapid, accurate, and robust tool for generating patient-specific 3D ossicle models from standard clinical CTs. This technology is expected to enhance the accuracy of biomechanical finite element simulations and serves as a foundational step toward advancing precision medicine in otologic surgery and custom prosthesis design.