📄 medicine

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

This single-center prospective study demonstrates that a novel machine learning framework, utilizing only a baseline CT scan combined with portable nasal airflow or respiratory inductive plethysmography signals, can achieve sub-millimeter real-time prediction of pulmonary nodule positions during free breathing, thereby enhancing the safety and efficiency of CT-guided lung punctures without requiring breath-holding or additional dynamic imaging.

Original authors: Yaqi Zeng, Tao Liu, Wen Tang, Dawei Wang, Xizhao Sui, Nan Hong, Libao Hu

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yaqi Zeng, Tao Liu, Wen Tang, Dawei Wang, Xizhao Sui, Nan Hong, Libao Hu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor trying to place a tiny needle into a small, moving target inside a patient's chest. This is a routine but delicate procedure used to diagnose or treat lung nodules, which are small growths that can be benign or cancerous. The challenge is that the lungs are never truly still; they expand and contract with every breath, causing the target to drift up and down. To compensate, doctors often ask patients to hold their breath, but this is difficult for many, especially the elderly or those with lung disease. If the patient cannot hold still, the doctor must take multiple X-ray scans to track the movement, which increases radiation exposure and prolongs the time the patient spends on the table. For years, the medical community has sought a way to predict exactly where that target will be at any given moment during a normal breath, allowing the doctor to aim accurately without forcing the patient to stop breathing or taking endless extra scans.

A team of researchers at Peking University People's Hospital has developed a new method to solve this problem using artificial intelligence and simple, portable sensors. In a recent study involving 139 patients, they tested a system that could predict the vertical movement of lung nodules with sub-millimeter precision. The researchers did not rely on expensive, complex machinery or new types of high-radiation scans. Instead, they used a standard planning CT scan to locate the nodule initially, and then attached lightweight, non-invasive sensors to the patients' faces and torsos. These sensors, similar to those used in sleep clinics, monitored the flow of air through the nose and the expansion of the chest and abdomen as the patients breathed naturally.

The core of the study was to see if a computer could learn the relationship between these breathing signals and the actual movement of the lung nodules. The team collected data from 143 nodules, recording the exact position of each nodule during multiple CT scans taken while the patients breathed freely. They fed this information into two different types of computer models: one based on a statistical method called a support vector machine, and another based on a deep learning algorithm known as ResNet-18. The goal was for the computer to look at the breathing signal at any moment and predict exactly where the nodule would be, even if the patient was in the middle of a breath cycle.

The results were remarkably accurate. When the models were tested on new data they had never seen before, the statistical error (measured as mean squared error) was approximately 0.166 to 0.198 mm². Crucially, the study explicitly confirmed that the actual absolute prediction errors for the nodule's position remained below 0.5 mm for all models. To put this in perspective, the thickness of a human hair is roughly 0.05 to 0.1 millimeters, meaning the prediction was precise enough to track movement far smaller than the width of a single hair. The system worked consistently well regardless of whether the patient was lying on their back, stomach, or side, and regardless of whether the nodule was located in the upper or lower part of the lung. The models performed slightly better when using the nasal airflow signal compared to the chest belt signal, but both methods achieved the high level of accuracy needed for clinical use.

This approach offers a significant shift in how these procedures could be performed. Currently, if a patient cannot hold their breath, the doctor must pause the procedure repeatedly to take new scans and check the needle's position, which adds time and radiation exposure. With this new predictive tool, the doctor could theoretically know where the nodule is at any second of the breathing cycle. This would allow for a smoother, faster procedure where the needle can be guided continuously without the need for frequent interruptions or forced breath-holding. The system requires only a standard CT scanner and a small, portable monitoring device, making it a low-cost solution that could be integrated into existing hospital workflows without the need for expensive robotic arms or major infrastructure changes.

The study was conducted as a single-center prospective trial, meaning the researchers recruited patients specifically for this study and followed them through the procedure, rather than looking back at old records. While the results are promising, the authors note that the system was tested primarily on patients undergoing a specific type of pre-surgery localization for lung nodules. They suggest that further testing is needed to confirm that the method works equally well for other procedures, such as taking tissue samples or burning away tumors, and across different hospitals with different equipment. However, the findings demonstrate that it is possible to track the invisible movement of a lung nodule in real-time using simple breathing signals and machine learning. By turning the chaotic motion of breathing into a predictable pattern, this technology could make lung interventions safer, faster, and less burdensome for patients who cannot hold their breath.

Technical Summary: Machine/Deep Learning-Based Position Prediction of Pulmonary Nodules in Free-Breathing CT-Guided Lung Puncture

Problem Statement
Respiratory motion, particularly cranio-caudal drift of pulmonary nodules driven by diaphragmatic movement, represents a primary obstacle to precision in free-breathing CT-guided thoracic interventions (e.g., microcoil localization, biopsy, and thermal ablation). This target drift necessitates frequent needle redirections and repeated CT acquisitions, leading to prolonged procedure times, increased cumulative radiation exposure for patients, and elevated risks of complications such as pneumothorax and pulmonary hemorrhage. Conventional mitigation strategies, including breath-holding and respiratory gating, are limited by high demands for patient compliance and restrict needle insertion to narrow, predefined respiratory phases, thereby reducing clinical utility, especially for elderly or compromised patients. While 4D-CT and robotic systems offer alternatives, they often involve higher costs, increased radiation, or operational constraints that limit their routine adoption.

Methodology
This single-center prospective cohort study enrolled 139 patients (143 nodules) undergoing preoperative CT-guided lung nodule localization for video-assisted thoracoscopic surgery (VATS). The study utilized a novel framework to predict nodule position in real-time without additional dynamic imaging.

Data Acquisition:
- Imaging: A single baseline planning CT was acquired. During the procedure, patients underwent standard CT scans (96-detector scanner) while breathing freely.
- Respiratory Signals: Continuous, non-invasive respiratory waveforms were captured using a portable sleep monitor (PSM) equipped with a nasal cannula (for nasal airflow) and thoracoabdominal belts (for Respiratory Inductive Plethysmography, RIP).
- Annotation: Nodule positions were defined as the Z-axis coordinates of the geometric center of the nodule's maximum cross-section. Signals were synchronized with CT scan timestamps.
- Preprocessing: Nasal airflow signals were zero-axis flipped to align with RIP peaks. RIP signals underwent sliding window baseline correction (6-second window) followed by Z-score normalization. Clinical metadata (age, sex, height, weight, lesion characteristics, and patient position) were also collected.
Model Architecture:
- Dataset: 2,275 paired CT datasets were split 4:1:1 for training, validation, and independent testing.
- Inputs: The models accepted patient clinical characteristics, initial nodule position, imaging features, and respiratory amplitude values corresponding to the initial and paired CT scans.
- Algorithms: Two approaches were compared:
  1. Support Vector Machine (SVM): Configured as SVR with a linear kernel.
  2. Deep Learning (ResNet-18): A dual-branch architecture where DICOM images were processed by ResNet-18 for image feature extraction, and non-image clinical features were processed by a ResNet-based extractor. These features were concatenated and fed into a final predictor.
- Output: The predicted Z-axis coordinate of the nodule in the subsequent respiratory phase.
- Metric: Performance was quantified using Mean Squared Error (MSE) in $mm^2$ .

Key Results
The study demonstrated that both SVM and Deep Learning models could achieve sub-millimeter prediction accuracy for nodule vertical displacement using only baseline CT and external respiratory signals.

Nasal Airflow Models:
- SVM: Achieved validation/test MSE of 0.099/0.166 $mm^2$ .
- ResNet-18: Achieved validation/test MSE of 0.120/0.186 $mm^2$ .
RIP Models:
- SVM: Achieved validation/test MSE of 0.100/0.168 $mm^2$ .
- ResNet-18: Achieved validation/test MSE of 0.122/0.198 $mm^2$ .
Error Magnitude: Across heterogeneous breathing patterns and lesion locations (including upper, middle, and lower lobes), the absolute prediction errors for central Z-coordinates remained below 0.5 mm for all constructed models.
Robustness: The models maintained performance across varying nodule sizes, densities (ground-glass, solid, sub-solid), and patient positions (supine, prone, lateral).

Key Contributions

Real-Time Prediction without Dynamic Imaging: The study validates a framework that predicts nodule position at arbitrary breathing phases using only a single baseline CT and low-cost, portable respiratory monitoring, eliminating the need for continuous dynamic CT scanning or 4D-CT.
Workflow Integration: Unlike phase-locked gating or breath-holding, this predictive system allows interventional radiologists to retain full flexibility in selecting needle insertion timing and trajectories without mandatory breath-hold coordination.
Hardware Accessibility: The solution relies on commercially available portable sleep monitoring hardware, avoiding the need for expensive robotic navigation systems or suite infrastructure renovations.
Comprehensive Feature Integration: The models successfully integrate clinical covariates (e.g., body habitus, position) and imaging features alongside respiratory signals to account for inter-patient variability in motion amplitude.

Significance and Claims
The authors claim that this approach directly addresses the clinical bottleneck of unpredictable respiratory target drift in free-breathing CT-guided interventions. By enabling sub-millimeter accuracy, the system aims to:

Reduce Radiation: Drastically lower the number of confirmatory CT scans required to verify needle-tumor alignment.
Improve Efficiency: Shorten total procedural runtime.
Enhance Safety: Mitigate risks of pneumothorax and pulmonary hemorrhage associated with repeated needle redirections and forced breath-holds.
Broad Applicability: Provide a technical foundation for next-generation robotic thoracic interventional platforms while offering immediate, low-cost improvements for standard manual CT-guided workflows.

The paper acknowledges limitations, including its single-center design, the exclusive focus on cranio-caudal (Z-axis) displacement without modeling anterior-posterior or left-right motion, and the derivation of data from preoperative localization cases rather than standalone biopsies or ablations. Future validation across diverse populations and scanner platforms is noted as necessary to confirm generalizability.

Technical Summary: Machine/Deep Learning-Based Position Prediction of Pulmonary Nodules in Free-Breathing CT-Guided Lung Puncture

More like this