Clinical Context Lung segmentectomy - surgical removal of a lung segment rather than an entire lobe - is increasingly preferred for early-stage non-small cell lung cancer because it preserves more lung function. However, precise preoperative 3D planning is essential due to the complex and variable anatomy of pulmonary vessels within each segment.
The Planning Bottleneck Traditional 3D surgical planning requires manual segmentation of pulmonary arteries, veins, bronchi, and the lung itself from CT scans - a process that takes trained specialists approximately 1.5 hours per patient. This time burden limits widespread adoption of anatomically guided segmentectomy planning.
AI Solution This study developed and validated an nnU-Net deep learning framework to automatically segment four key pulmonary structures from CT scans: pulmonary arteries, pulmonary veins, bronchi, and lung parenchyma. Automation reduces planning time to under 5 minutes while maintaining high accuracy.
Clinical Validation Beyond technical performance, the AI-generated segmentations were used in actual robotic-assisted thoracoscopic surgery (RATS) segmentectomy procedures using PulmoSR surgical planning software, validating real-world utility in the operating room.
Dataset The study used 125 CT scans: 100 for training and validation, and 25 for testing. Each scan was manually annotated by expert radiologists to create ground truth segmentation masks for all four target structures, providing the supervised learning signal for model training.
nnU-Net Framework nnU-Net (no-new-Net) is a self-configuring deep learning segmentation framework that automatically adapts its architecture, preprocessing, and training strategy to the specific dataset. This reduces manual hyperparameter tuning and is state-of-the-art across many medical image segmentation benchmarks.
Four Separate Models Individual nnU-Net models were trained for each of the four target structures: pulmonary artery, pulmonary vein, bronchus, and lung parenchyma. Separate models allow each structure's unique characteristics - such as the thin walls of bronchi versus the solid parenchyma - to be optimized independently.
Evaluation Metrics Performance was measured using Dice similarity coefficient (overlap accuracy), sensitivity (detection completeness), and specificity (false positive rate). These metrics together assess both the precision and recall of the segmentation, which are critical for surgical planning accuracy.
Dice Scores All four nnU-Net models achieved Dice similarity coefficients of 0.91 to 0.92, indicating high spatial overlap between AI-generated and manually annotated segmentations. This level of accuracy is generally considered sufficient for clinical use in surgical planning applications.
Sensitivity and Specificity Sensitivity ranged from 0.84 to 0.86, meaning the models correctly identified 84 to 86 percent of all true vessel and airway voxels. Specificity was 0.99 for all structures, confirming an extremely low false positive rate and minimal inclusion of non-target tissue.
Consistency Across Structures The similarity in performance across all four anatomical structures - despite their different sizes, shapes, and imaging characteristics - demonstrates the robustness and generalizability of the nnU-Net approach for comprehensive thoracic anatomy segmentation.
Time Efficiency The most clinically impactful result was the reduction in planning time from approximately 1.5 hours with manual segmentation to under 5 minutes with the AI pipeline. This 18-fold speedup makes routine 3D planning feasible even in high-volume centers.
Surgical Integration AI-generated segmentations were integrated with PulmoSR, a commercial surgical planning software, and used in five robotic-assisted thoracoscopic segmentectomy procedures. This clinical use case demonstrates the full pipeline from CT acquisition to operating room deployment.
Procedural Success All five procedures using AI-derived 3D anatomical maps were completed successfully. Surgeons reported that the reconstructions accurately reflected intraoperative anatomy, providing reliable guidance for identifying vessel and bronchus locations during the procedure.
Safety Considerations The combination of high specificity (0.99) with adequate sensitivity means the models rarely generate false vessel representations that could mislead surgeons. For safety-critical surgical guidance, minimizing false positives is especially important.
Comparison to Manual Planning While the study did not include a randomized comparison with manual planning in surgical outcomes, the qualitative concordance between AI maps and intraoperative findings in all five cases supports the potential equivalence of AI-guided planning to expert-crafted reconstructions.
Scaling Surgical Access The most immediate impact of this technology is democratization of 3D surgical planning. If centers without dedicated CT segmentation specialists can generate accurate 3D reconstructions in minutes, anatomically precise segmentectomy becomes accessible to a much wider patient population.
Integration with Intraoperative Guidance Future work could explore real-time integration of AI-generated 3D maps with intraoperative imaging or augmented reality systems, allowing surgeons to overlay preoperative reconstructions onto the live surgical field for improved navigation.
Larger Validation Studies The current validation in 25 test scans and 5 surgeries represents a promising proof of concept but is limited in scale. Multicenter studies with hundreds of cases are needed to confirm generalizability across scanner types, patient anatomies, and different surgical centers.
Extension to Other Procedures The same AI segmentation pipeline could potentially support planning for other thoracic procedures including lobectomy, sleeve resection, and bronchoplasty, as well as ablation planning and radiation therapy target delineation.