Deep Learning-based Framework for Rapid, Precise and Fully Automated Localization of Cochlea Landmarks in 3D CT Images
Year: 2026
Type: preprint
Abstract: Manual localization of anatomical landmarks within human cochlea structures in 3D computed tomography (CT) images is labor-intensive. However, despite progress in various studies on automatic localization methods, it remains challenging to jointly achieve rapid, precise, and fully automated landmark localization. Therefore, we propose a deep learning-based localization framework, called MEDICO-YOLO, which bridges speed and precision. It can automatically and simultaneously localize 6 human cochlea landmarks (apex, basal point, round window center of both left and right cochleae) in the full-resolution 3D CT images. The main contributions are: 1) MEDICO-YOLO learns from few manually annotated 3D CT images; 2) Within the framework, a 2D CNN-based network was developed to specifically localize cochlea landmarks in 3D single-scale medical CT images; 3) A novel loss function was designed; 4) The localization model achieved human expert-level precision while ensuring rapid inference. This study employed anonymized 3D pre-operative CT images (0.3 mm, covering the full head or temporal bone region). The evaluation involved comparing four models that were trained using 4 different training set sizes (n=4, 8, 12, 16). Each of them was evaluated on a distinct test set (n=16). The results show that the MEDICO-YOLO framework achieved strong localization performance by learning from few annotated 3D CT images. Trained on only 8 CT volumes, the model already attained a mean localization error of 1.30 mm (approximately 4 voxels). When trained on 16 CT volumes, the model achieved a mean error of 0.71 mm and a successful detection rate ([email protected] mm) of 97.92%. Furthermore, on an Apple M1 laptop, the model achieved a real-time speed (32 ms) for inferencing each 2D slice (608 x 608) and an average latency of 25.43 s for inferencing the full 3D CT image (608 x 608 x 200, including post-processing). The complete methodology for the MEDICO-YOLO framework is presented in this article.
Source: medRxiv
Authors: Yifan Wang, Andrej Kral, Thomas Lenarz, Samuel John
Institutions: Medizinische Hochschule Hannover, Hearing4all
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: European Commission
OpenAlex ID: W7225559489