Deep Learning-Based 3D Segmentation and Reconstruction of Laryngeal Anatomy from MRI

This is the pre-peer reviewed version of the article submitted to The Laryngoscope for peer review. This article may be used for non-commercial purposes in accordance with Wiley Terms and Conditions for Use of Self-Archived Versions. Abstract Objective: To develop and evaluate a deep convolutional neural network for automatic threedimensional reconstruction of laryngeal cartilages and the glottal airway from MRI. Methods: High-resolution ex vivo MRI scans were acquired from 16 harvested rabbit larynges. Manual segmentations of the thyroid cartilage, cricoid cartilage, arytenoid cartilage, and glottal 2 airway were created using ITK-SNAP and refined by expert raters. A three-dimensional nnU-Net framework was trained using five-fold cross-validation. Segmentation performance was assessed using the Dice Similarity Coefficient (DSC), Hausdorff distance (HD), and average surface distance (ASD). A blinded observer study was conducted in which two expert raters independently evaluated manual and automated segmentations on a 4-point accuracy scale, with interobserver agreement quantified using the Intraclass Correlation Coefficient (ICC). Results: The model achieved high volumetric agreement across all structures, with mean DSC values exceeding 0.80. Cartilaginous structures demonstrated higher segmentation stability than the glottal airway, with lower HD and ASD values. Performance distributions were consistent across cases, indicating robust anatomical generalization. Automated segmentations demonstrated substantially higher interobserver agreement than manual segmentations (ICC(2,1) = 0.841 vs. 0.339) and produced smoother three-dimensional surface reconstructions with fewer boundary irregularities. Conclusion: Deep learning-based segmentation enables reliable three-dimensional reconstruction of laryngeal anatomy from MRI, supporting patient-specific modeling and surgical planning. Automated segmentations were more consistent across raters and produced higher-quality surfaces than manual delineations, supporting patient-specific modeling and surgical planning.