An automated deep learning-based pipeline for 3D characterisation of the murine upper airway from micro-CT.

Introduction: Upper airway structure and function are affected in several respiratory (e.g., obstructive sleep apnoea, allergic rhinitis, and asthma) and congenital [e.g., Down syndrome (DS)] conditions. Although murine models are widely used to study airway pathology and to evaluate therapies, quantitative analysis of the upper airway on micro-computed tomography (micro-CT) remains limited as it is hampered by labour-intensive and observer-dependent manual segmentation of the images. Automated and standardized image analysis methods are therefore needed to support mechanistic and translational upper airway research. Materials and methods: We developed a deep learning-based pipeline for fully automated segmentation and characterization of the trachea, pharynx, and nasal cavity from micro-CT images. A full-resolution 2D CNN based on the nnU-Net framework was trained in the axial plane using 5-fold cross-validation on 106 scans from a DS study involving wild-type (WT) and trisomic (TS) mice. Generalisability was assessed on an external validation set of 74 scans from another DS study examining the effect of RSV infection in WT and TS mice. Model performance was evaluated using the Dice similarity coefficient (DSC), relative absolute volume difference (RAVD), Pearson correlation and Bland-Altman analysis between manual and automated measurements. Results: Strong segmentation performance was obtained on the internal validation set, with DSC of 0.988 ± 0.010, 0.985 ± 0.008, and 0.958 ± 0.004, and RAVD of 1.483% ± 1.800%, 1.506% ± 1.217% and 1.965% ± 1.563% for trachea, pharynx and nasal cavity, respectively. Performance remained robust on the external validation set, confirming generalisability across different experimental conditions and imaging protocols. Automatically extracted volumetric biomarkers showed strong agreement with manual measurements and small Bland-Altman biases. The pipeline reproduced biologically relevant group differences previously identified through manual analysis, including reduced pharynx and nasal cavity volumes in TS mice compared with WT controls. Conclusions: We present a deep learning-based pipeline for automated, standardized 3D characterization of the murine upper airway from micro-CT. By reducing manual workload and enhancing measurement consistency, it enables high-throughput phenotyping and mechanistic investigation of the upper airway in preclinical models. By facilitating direct, automated measurements from imaging, our pipeline strengthens the translational bridge between experimental and clinical research.