Abstract Background Micro-computed tomography (micro-CT) imaging is essential for assessment of lung pathology in preclinical fibrotic respiratory disease models. However, manual lung tissue segmentation is time-intensive, requires specialized expertise, and is associated with substantial observer variability. Existing automated segmentation methods lack generalizability across different scanning protocols and imaging conditions, and are particularly poor at segmenting fibrotic lungs. We developed and validated a novel domain-adaptive 3D deep learning framework designed to overcome these limitations and provide robust automated lung segmentation from mouse micro-CT images. Methods We analyzed micro-CT scans of normal and bleomycin-treated lungs from 44 mice, comprising 156 total scans acquired across multiple micro-CT scanners and timepoints at the University of Wisconsin-Madison. The dataset was partitioned into training (127 scans, 35 mice, 80%) and validation (29 scans, 9 mice, 20%) cohorts. Our Domain-Adaptive 3D U-Net architecture integrates three synergistic components: (1) adaptive intensity normalization using multi-scale convolutional analysis to dynamically adjust tissue-specific transformation parameters; (2) hierarchical domain adaptation modules with domain-specific instance normalization layers integrated at multiple network depths to learn representations that are robust to domain shifts; and (3) shape-aware guidance that learns both local tissue details and global anatomical context to enforce anatomical plausibility through spatial attention maps. Strategic patch-based training was employed to address the computational challenge of processing large 3D volumes by utilizing 300 randomly selected patches from each scan during the training phase, and all non-overlapping patches during the validation phase. Performance was subsequently compared to standard U-Net, Attention U-Net, and Dynamic U-Net architectures, each trained using a similar pipeline and evaluated using the Dice Similarity Coefficient (DSC). Results The Domain-Adaptive U-Net demonstrated performance gains over existing architectures in bleomycin-injured lungs in training and validation datasets (Figure 1). During training, our model achieved a DSC of 0.91, significantly outperforming standard U-Net (DSC=0.87, +4.1% improvement), Attention U-Net (DSC=0.81, +10.4% improvement), and Dynamic U-Net (DSC=0.89, +2.6% improvement). The performance gain was more pronounced during validation, where our framework achieved a DSC of 0.95, the highest among all tested architectures. This represented substantial improvements over standard U-Net (DSC=0.89, +6.7% improvement), Attention U-Net (DSC=0.88, +7.8% improvement), and Dynamic U-Net (DSC=0.80, +15.6% improvement). Qualitative analysis revealed superior performance in anatomically challenging regions, including fibrotic tissue and irregular lung boundaries caused by pathological changes. Conclusion Our proposed domain-adaptive deep learning framework provides an efficient, automated solution for mouse lung segmentation from micro-CT images, demonstrating excellent accuracy essential for high-throughput preclinical respiratory research. This abstract is funded by: NIH/NHLBI (K01HL163249, R01HL151421, HL171464), American Heart Association (24CDA1272732), Pulmonary Fibrosis Foundation (1053163)
Harr et al. (Fri,) studied this question.