Het versterken van experimenteel en klinisch longonderzoek door AI-gestuurde kwantitatieve CT analyse

Respiratory diseases such as infections, fibrosis, and cancer remain leading causes of death worldwide. Despite extensive research into new therapeutic strategies, many drug candidates identified in animal models failed in clinical trials, highlighting a translational gap between experimental and clinical pulmonary research. Beyond the well-recognized limitation that experimental models cannot fully capture the complexity of human disease, this translational problem raises the question of whether the methodological toolbox used throughout the experimental and clinical research phases should be expanded and better aligned. In this thesis, we proposed integrating quantitative computed tomography (CT) of the lung into both experimental and clinical trials as a crucial complement to existing outcome measures, thereby establishing a common language across both trial types. Because manual analysis of high-dimensional imaging data is resource-intensive and prone to reader variability, artificial intelligence (AI) techniques such as deep learning (DL) using convolutional neural networks (CNN), are increasingly used for the quantification of CT, though their application to micro-CT is still limited. The goal of this PhD thesis was to implement AI-powered quantitative (micro-)CT in both experimental and clinical settings to improve insights into disease progression and treatment response, which could ultimately enhance translational knowledge exchange and help bridge the gap between experimental findings and clinical outcomes. In the first study, we addressed a major bottleneck in experimental respiratory imaging research by developing a generic DL-based pipeline for automated quantification of lung pathology from micro-CT. Using diverse longitudinal micro-CT datasets spanning multiple disease models, including genetic disorders and viral and fungal infections, with variable lung pathology and degree of disease burden, we trained 2D CNNs on axial, coronal and sagittal slices and combined these single-orientation models into a 2.5D model. The final 2.5D model achieved excellent lung segmentation performance and showed strong generalisability across lung pathologies, acquisition protocols and rodent species. Moreover, we proved that our proposed solution effectively monitors longitudinal lung pathology development and response to treatment in real-world preclinical studies. In the second study, we expanded our suite of automated micro-CT quantification tools to the murine upper airway, which is studied in experimental models of several upper airway-affecting respiratory (e.g. obstructive sleep apnoea, allergic rhinitis, and asthma) and congenital (e.g. Down syndrome (DS)) conditions. We developed a DL-based pipeline for automated segmentation and quantitative characterization of the trachea, pharynx and nasal cavity. Using datasets from DS mouse models acquired under varying experimental conditions, we trained a high-resolution 2D CNN. The model demonstrated excellent segmentation performance and generalisability across independent studies and imaging protocols, and successfully reproduced biologically relevant group differences previously identified through manual analysis. Together with the first study, this method expands our toolkit for efficient, standardized, high-throughput quantification of respiratory structures from micro-CT, promoting translational experimental research using imaging. In the third study, we moved to the clinical setting to investigate whether automated CT analysis could aid in the detection and quantification of interstitial lung disease. We developed a DL-based model for the Automated Recognition of Interstitial Lung Abnormalities (ARIA) on clinical CT, designed to detect and quantify key radiological lung patterns, including reticulation, ground-glass opacity, consolidation, and honeycombing, across diverse imaging protocols. Trained and validated on carefully annotated datasets, ARIA demonstrated strong detection and quantification performance in agreement to manual assessments, while remaining robust to protocol variability. In a clinical cohort of patients with systemic sclerosis-associated interstitial lung disease, ARIA-derived pattern volumes differentiated progressive from non-progressive disease and correlated with pulmonary function tests, highlighting clinical relevance. By enabling standardized, objective quantification of subtle disease patterns on CT, this approach supports more precise characterization and monitoring of interstitial lung disease. In the fourth study, we investigated whether automated quantitative CT analysis could support the detection of symptomatic pneumonitis in patients with stage III unresectable non-small cell lung cancer treated with chemoradiotherapy with or without durvalumab immunotherapy. Using imaging and clinical data from the PACIFIC clinical trial, we developed and internally validated logistic regression models based on clinical variables, quantitative CT-derived measurements including ARIA-derived pattern volumes, and a combination of both. While clinical and imaging models individually showed good discriminative performance, the combined model performed best and identified key risk factors spanning demographics, treatment variables and imaging variables, including normalized consolidation, reticulation and lung volumes. All models demonstrated high negative predictive value, indicating potential utility for symptomatic pneumonitis screening. These findings show that integrating automated quantitative imaging with clinical data can enhance detection of treatment-related lung toxicity and further illustrate the clinical utility of standardized CT-based measurements. In conclusion, this thesis established imaging, particularly quantitative CT, as a powerful translational technology that complements traditional outcome measures by enabling direct, spatially and temporally resolved assessment of complex pulmonary responses. Across four use cases, we demonstrated that AI-powered imaging biomarkers provide complementary insights into disease progression and treatment response in both preclinical and clinical settings. By developing and validating deep learning models for automated image processing, we addressed key barriers to large-scale adoption of quantitative imaging, providing reliable, standardized measurements that strengthen the translational link between experimental and clinical studies. Collectively, this thesis advanced the concept of imaging as a common language across research settings, demonstrated the (pre-)clinical relevance of automatically derived imaging biomarkers, and ensured the sustained use of the developed tools and models beyond this PhD.