Synchrotron-based deep learning for intracochlear segmentation of clinical scans with international validation

Clinical imaging is routinely used pre-operatively for cochlear implantation, yet lacks the resolution and contrast necessary to visualize the fine intracochlear structures critical for individualized intervention. To address this limitation, an ensemble deep learning model was developed to automatically segment cochlear micro-anatomy from standard clinical scans. The model was trained and validated using an independent internal dataset comprised of paired synchrotron and clinical scans of the same cochleae across various acquisition protocols. Performance was evaluated quantitatively on an unseen internal test dataset and a multi-institutional external test dataset. The deep learning model achieves accurate segmentation of the scala tympani (ST) and scala vestibuli (SV) across all tested modalities, with a mean Dice similarity coefficient of 0.895 ± 0.024 and 0.891 ± 0.029, a max Hausdorff distance of 0.396 ± 0.100 mm and 0.444 ± 0.280 mm, and an average Hausdorff distance of 0.008 ± 0.003 mm and 0.008 ± 0.003 mm, respectively. The model achieves performance metrics superior to previously published scalar models and demonstrates strong viability on the multi-institutional external dataset. Furthermore, anatomical measurements on the automatic segmentations closely match those obtained from high-resolution ground truth segmentations, measured using scalar volume and lateral ST length, confirming reliable estimation of clinically relevant metrics. By bridging the gap between high-resolution imaging and routine clinical imaging, this work could provide a practical solution for patient-specific cochlear implant surgical planning through electrode selection and post-operative assessment through image fusion, advancing the goals of atraumatic insertions and more effective hearing restoration. Micuda et al. develops and validates an ensemble deep learning model that segments cochlear micro-anatomy using paired synchrotron and routine clinical scans. The model accurately delineates the scala tympani and scala vestibuli across imaging configurations, outperforming prior methods and preserving clinically relevant measurements. Imaging used before cochlear implant surgery often does not have enough detail to visualize the small structures inside the cochlea that are important for surgical planning and patient outcomes. To address this limitation, a deep learning model was developed to automatically extract the small anatomical structures from routine clinical scans. The model was developed using paired high-resolution and clinical scans of cadaveric cochleae. It was subsequently evaluated quantitatively on a cadaveric dataset, assessed for clinical plausibility on an international external patient dataset, and compared against previously published models. The model achieved high accuracy on both objective and cochlear-specific metrics across all imaging configurations, closely matched the high-resolution reference images, and outperformed similar models reported in the literature. This deep learning model has the potential to be used in clinical practice to improve surgical planning for cochlear implantation, support patient-specific implant selection, and aid in the assessment of the implant position after surgery.