Abstract
Idiopathic pulmonary fibrosis (IPF), the archetype of pulmonary fibrosis (PF), is a chronic lung disease of a poor prognosis, characterized by progressively worsening of lung function. Although histology is still the gold standard for PF assessment in preclinical practice, histological data typically involve less than 1% of total lung volume and are not amenable to longitudinal studies. A miniaturized version of computed tomography (µCT) has been introduced to radiologically examine lung in preclinical murine models of PF. The linear relationship between X-ray attenuation and tissue density allows lung densitometry on total lung volume. However, the huge density changes caused by PF usually require manual segmentation by trained operators, limiting µCT deployment in preclinical routine. Deep learning approaches have achieved state-of-the-art performance in medical image segmentation. In this work, we propose a fully automated deep learning approach to segment right and left lung on µCT imaging and subsequently derive lung densitometry. Our pipeline first employs a convolutional network (CNN) for pre-processing at low-resolution and then a 2.5D CNN for higher-resolution segmentation, combining computational advantage of 2D and ability to address 3D spatial coherence without compromising accuracy. Finally, lungs are divided into compartments based on air content assessed by density. We validated this pipeline on 72 mice with different grades of PF, achieving a Dice score of 0.967 on test set. Our tests demonstrate that this automated tool allows for rapid and comprehensive analysis of µCT scans of PF murine models, thus laying the ground for its wider exploitation in preclinical settings.
Introduction
Preclinical research into Idiopathic Pulmonary Fibrosis (IPF) is essential for developing new therapies, yet it faces a significant methodological bottleneck. The current "gold standard," histology, requires animal sacrifice, which prevents longitudinal studies of the same subject. Furthermore, histology only samples approximately 1% of the total lung volume, potentially missing localized fibrotic changes.
Micro-computed tomography (µCT) offers a non-invasive alternative that captures the entire lung volume. By measuring X-ray attenuation, researchers can calculate "densitometry" to map the ratio of air to tissue. However, fibrosis changes lung density so drastically that standard automated tools often fail, forcing researchers into a grueling 40-minute manual segmentation process per scan. This study introduces a deep learning solution to automate this workflow.
Methods
The proposed pipeline utilizes a "coarse-to-fine" strategy to balance computational efficiency with high-resolution accuracy. The process begins with a 2D U-Net operating on low-resolution axial slices to identify the lungs, heart, and airways. This stage also performs an "internal phantom" calibration: by identifying the heart and airways, the system automatically converts raw gray levels into Hounsfield Units (HU), a standard scale for radiodensity, without needing external reference objects.
The core segmentation is handled by a 2.5D CNN approach. Instead of a single, memory-heavy 3D model, the pipeline uses three separate 2D U-Nets trained on orthogonal planes (axial, sagittal, and coronal). The predictions are integrated via majority voting, capturing 3D spatial coherence while maintaining high resolution. Finally, the segmented lungs are compartmentalized into functional zones (e.g., normo-aerated vs. non-aerated) based on density thresholds.
Key Results
The pipeline was validated on 219 µCT scans of female mice, including both healthy controls and bleomycin-induced fibrosis models. Key findings include:
- High Accuracy: The model achieved a Dice score of 0.967, indicating near-perfect overlap with expert manual segmentation.
- Significant Speedup: The automated process takes approximately 81 seconds per scan, compared to over 40 minutes for manual analysis—a 30x improvement in throughput.
- Robustness: The correlation between automated and manual densitometry was exceptionally high ($R^2 = 0.99$). The model also generalized well to different breathing phases (end-inspiration vs. end-expiration), maintaining a Dice score of ~0.95.
- Transfer Learning: By first training the network to recognize "lungs" and then refining it to distinguish between left and right sides, the authors successfully handled datasets with limited labels.
Conclusion
This fully automated deep learning pipeline effectively removes the manual segmentation bottleneck in preclinical lung research. By providing rapid, accurate, and non-invasive densitometry, it enables high-throughput longitudinal studies that were previously impractical. This tool is poised to accelerate pharmaceutical drug discovery and improve the assessment of pulmonary fibrosis in murine models.
Summary updated