How Deep Learning Is Changing Lung Imaging
Summary
- A new study from the University of Miami Miller School of Medicine is using artificial intelligence to identify lung segments from CT scans with near-radiologist-level accuracy.
- The Miller School research team built a multi-stage, 3D convolutional neural network (CNN) that mimics how radiologists trace airways and identify bronchi.
- The model was tested on 20 independent CT scans and reviewed by radiologists, achieving a 99.2% match rate with expert-defined segments.
What if physicians could map the lungs down to their smallest functional units without ever stepping into an operating room?
A new study from the University of Miami Miller School of Medicine in collaboration with Siemens Medical Solutions is using artificial intelligence to develop a deep-learning algorithm that can automatically identify lung segments from CT scans with near-radiologist-level accuracy.
“Assessing segmental variation in ventilation and perfusion can assist with early detection of disease as well as response to therapeutics,” said co-first author Trishul Siddharthan, M.D., an associate professor in the Miller School’s Division of Pulmonary, Critical Care and Sleep Medicine. “Respiratory disease is expected to double over the next decade. There is a need to detect disease and determine which patients benefit from novel therapeutics and procedures.”
Published in the Journal of Applied Clinical Medical Physics, this research is changing how providers diagnose and treat respiratory diseases like chronic obstructive pulmonary disease (COPD). The research, in essence, is a glimpse into the future of radiology, pulmonology and precision medicine.
The Challenge of Mapping Lung Segments
Lung segments have their own bronchus and artery and are functionally independent, which means clinicians can target them for surgery, therapy or imaging without affecting surrounding tissue. But on a standard CT scan, these segment boundaries aren’t clearly visible. That’s where AI steps in.
The Miller School research team built a multi-stage, 3D convolutional neural network (CNN) that mimics how radiologists trace airways and identify bronchi. Here’s how it works:
• Airway tree extraction: The algorithm starts by identifying the centerlines of the airway tree using a CNN trained on probability maps. It uses rivulet tracing—a computational technique used to reconstruct branching structures by iteratively backtracking through 3D image data to ensure continuity even in noisy or distorted regions.
• Bronchial labeling: Using the open-source tool ParaView, annotators label each of the 18 segmental bronchi. These labels are propagated through the airway branches, dramatically reducing manual effort.
• Segment mask generation: The algorithm calculates the shortest distance from each voxel in the lung to its nearest segmental bronchus. This spatial relationship is used to assign segment labels, creating a full, 3D map of the lung’s internal architecture.
• End-to-end deep learning: The team trained a deep image-to-image (DI2I) network on 123 annotated CT scans. During inference, the model uses deep reinforcement learning to locate anatomical landmarks like the carina bifurcation and sternum tips, ensuring accurate cropping and segmentation.
• Validation: The model was tested on 20 independent CT scans and reviewed by radiologists. It achieved a 99.2% match rate with expert-defined segments, even in patients with COPD, where lung anatomy is often distorted.
“Standard lung imaging modalities do not allow us to determine dynamic changes while we are breathing,” said Dr. Siddhartan. “Prior functional assessments like single positron emission tomography and functional MRI give us rough estimates of imbalances in ventilation and perfusion across large regions of the lung. But they do not provide precise estimates at the anatomic level of the airways or lung segments. This technology can allow us to determine changes in regional ventilation at the levels of the airways that opens the possibility of early detection of disease as well as early response to therapy.”
Clinical Impact: Precision Medicine in Practice
Dr. Siddharthan and team have effectively used innovative technology to create a tool with real clinical power. Its clinical advantages include:
• Lung volume reduction: In COPD treatment, segment-level imaging helps pinpoint poorly ventilated areas for targeted removal, improving breathing and preserving healthy tissue.
• Targeted drug delivery: Inhaled medications don’t spread evenly. Segment-level maps can guide where drugs are most likely to work, especially for biologics targeting airway remodeling.
• Functional imaging fusion: Pairing this AI with PET or SPECT scans allows clinicians to overlay ventilation and perfusion data onto anatomical segments, enhancing early detection and treatment monitoring.
• Bronchoscopic navigation: Whether placing valves or performing biopsies, having a detailed segment map improves accuracy and safety.
• Post-surgical planning: Segment identification supports better predictions of lung function after surgery and helps tailor recovery plans.
What’s Next? Future Research Directions
While this study focused on COPD, the potential applications are much broader:
• Expanding to other lung diseases: Conditions like interstitial lung disease and lung cancer often distort airway anatomy. Future studies will test the algorithm’s accuracy in these more complex scenarios.
• Larger and more diverse cohorts: To improve generalizability, researchers plan to train and validate the model on larger datasets that include a wider range of ages, disease types and imaging conditions.
• Comparing modalities: The team aims to compare their CNN-based approach with atlas-based and hybrid segmentation techniques to determine which method performs best across different clinical settings.
• Integrating functional data: By combining AI-generated segment maps with functional imaging (like SPECT or PET), future tools could provide a real-time view of how each lung segment contributes to breathing and open doors to personalized treatment strategies.
• Improving training efficiency: Annotating CT scans is time-consuming. Researchers are exploring semi-supervised learning and active learning techniques to reduce the need for manual labeling while maintaining accuracy.
Tags: AI, artificial intelligence, chronic obstructive pulmonary disease, COPD, CT scan, Division of Pulmonary, Critical Care and Sleep Medicine, Dr. Trishul Siddharthan, lung cancer, lung imaging, pulmonary medicine