Towards fully automated Inner Ear Analysis: Deep-Learning-based Joint Segmentation and Landmark Detection Framework

Automated analysis of the inner ear anatomy in radiological data instead of time-consuming manual assessment is a worthwhile goal that could facilitate preoperative planning and clinical research. We propose a framework encompassing joint semantic segmentation and anatomical landmark detection. A fully automated pipeline with a single, dual-headed volumetric 3D U-Net was implemented, trained and evaluated using manually labeled in-house datasets from cadaveric specimen (N = 44) and clinical practice (N = 10). The model robustness was further evaluated on two independent open-source datasets (N = 23 + 7 scans). For the in-house datasets, Dice scores of 0.965 and 0.9505, intersection-over-union scores of 0.934 and 0.906 and average Hausdorff distances of 0.085 and 0.096 voxel units were achieved. The landmark localization task was performed automatically with an average localization error of 5.36 and 5.77 voxel units. A robust, but reduced performance could be attained for the two open-source datasets. The feasibility of the suggested integrated segmentation and landmark detection framework was experimentally evident by achieving competitive, efficient and consistent performance across a multitude of realistic, clinically relevant datasets and evaluation metrics. Based on these results, the applications in clinical practice and research beyond the data selection utilized in this study can be accomplishable.