Machine/Deep Learning-Based Position Prediction of Pulmonary Nodules in Free- Breathing CT-Guided Lung Puncture: A Single-Center Prospective Cohort Study

Abstract Background Cranio-caudal respiratory drift of pulmonary nodules severely compromises accuracy of free-breathing CT-guided thoracic interventions, contributing to repeated CT acquisitions, prolonged procedure time, elevated patient radiation burden, and higher complication risks of pneumothorax and pulmonary hemorrhage. Conventional motion mitigation approaches such as breath-holding and respiratory gating require high-level patient compliance and constrain needle insertion to narrow predefined respiratory phases, which limits real-world clinical utility. This work addresses this unmet clinical need by developing and validating novel real-time artificial-intelligence prediction models. Leveraging only one baseline planning CT together with low-cost, portable nasal airflow and respiratory inductive plethysmography (RIP) signals, our framework achieves sub-millimeter prediction of nodule cranio-caudal position without additional dynamic imaging. Methods This single-center prospective cohort enrolled 139 patients with 143 nodules undergoing pre-VATS CT microcoil localization. Synchronized respiratory waveforms, paired CT scans, clinical, lesion and procedural metadata were collected. Nodule vertical displacement was quantified via Z-axis coordinate shifts of each nodule’s maximum cross-section geometric center. In total, 2275 paired CT datasets were split 4:1:1 for model training, validation and independent testing. Prediction performance was quantified using mean squared error (MSE). Results The nasal-airflow-driven SVM yielded validation/test MSE values of 0.099/0.166 mm², versus 0.120/0.186 mm² for ResNet-18. RIP-based SVM achieved validation/test MSE of 0.100/0.168 mm² compared with 0.122/0.198 mm² for ResNet-18. Across heterogeneous breathing patterns and lesion locations, absolute prediction errors of central Z-coordinates remained below 0.5 mm for all constructed models. Conclusions Portable respiratory monitoring coupled with machine-learning enables sub-millimeter nodule Z-position prediction without mandatory breath-hold or phase-locked gating. This solution can be directly integrated into existing manual CT-guided puncture workflows to improve procedural safety and efficiency with minimal additional hardware; it also provides an essential technical building block for next-generation robotic thoracic interventional platforms. Trial registration: ChiCTR2300068696 (2023-02-27)