Deep Learning-based Framework for Rapid, Precise and Fully Automated Localization of Cochlea Landmarks in 3D CT Images

MEDICO-YOLO localizes 6 cochlea landmarks in full-resolution 3D CT, reaching a mean error of 0.71 mm and SDR@3.0 mm of 97.92% when trained on 16 volumes.

• Why it matters: Manual 3D cochlea landmarking is labor-intensive, and prior automatic methods struggle to combine speed, precision, and full automation in one workflow.
• What they did: The framework uses few-shot learning from anonymized pre-operative 3D CT (0.3 mm) and compares four models trained on n=4, 8, 12, and 16 volumes, each tested on n=16.
• The result: With only 8 CT volumes, mean localization error is 1.30 mm (~4 voxels), and on an Apple M1 laptop inference takes 32 ms per 2D slice and 25.43 s per full 3D CT (608×608×200).

AI summary of the abstract · may contain errors · not medical advice · read the paper.