Airway image analysis systems: a review of datasets, evaluation methodologies, and deep learning approaches

Airway image analysis systems play an increasingly important role in computer-assisted diagnosis, quantitative airway assessment, intervention planning, and respiratory disease management. However, the development of robust airway analysis methods remains challenging due to the complex geometry of airway structures, strong class imbalance, variability in imaging protocols, and the need to preserve anatomical connectivity across multiple spatial scales. Recent advances in topology-aware learning, transformer-based architectures, and foundation-model adaptation have substantially changed the landscape of airway image analysis, motivating an updated review focused specifically on airway-oriented systems. In this narrative review, we provide a comprehensive overview of the datasets, evaluation methodologies, and computational approaches that underpin modern airway image analysis systems. We discuss the imaging modalities and publicly available datasets commonly used for airway analysis and examine their strengths, limitations, and implications for algorithm development and benchmarking. Particular attention is given to evaluation methodologies, including overlap-based, surface-distance, and topology-aware metrics that assess both segmentation accuracy and structural correctness. We further review the evolution of computational approaches, covering convolutional neural networks, transformer-based architectures, hybrid methods, topology-aware and tubular-aware networks, foundation-model adaptations, and data-efficient learning strategies. Finally, we discuss current challenges related to dataset availability, annotation quality, model generalization, topology preservation, and clinical deployment, and outline future research directions toward robust, interpretable, and clinically deployable airway image analysis systems.