Deep Learning-Based Lung Segmentation for Robotic Electromagnetic Navigation in the lungs
Principal Investigator
Name
Ron Barak
Degrees
B.Sc. Mathematics
Institution
Magnisity Ltd.
Position Title
CTO & Co-founder
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1436
Initial CDAS Request Approval
Jul 11, 2025
Title
Deep Learning-Based Lung Segmentation for Robotic Electromagnetic Navigation in the lungs
Summary
We aim to develop and validate deep learning models for automatic lung segmentation using chest CT scans from the National Lung Screening Trial (NLST) dataset. These models will be used as a key component in a robotic electromagnetic navigation system for image-guided thoracic procedures.
Robotic navigation in the lungs requires accurate registration between the patient’s anatomy and the preoperative CT scan. To achieve this, we rely on robust segmentation of lung structures, which is the first and most critical step in the registration pipeline. While small internal datasets exist, they are insufficient to train generalizable deep neural networks.
The NLST dataset offers a unique opportunity due to its large size, diversity of imaging conditions, and focus on low-dose chest CT. We plan to use only the CT image data, without any access to patient identifiers or clinical outcomes. The segmented images will be used to train and evaluate models that can perform with high accuracy and robustness across varied anatomies.
The final models will be integrated into a clinical software system under development, intended for use in interventional pulmonology. The project aligns with the broader goal of improving the precision and safety of minimally invasive procedures using AI and robotics.
Robotic navigation in the lungs requires accurate registration between the patient’s anatomy and the preoperative CT scan. To achieve this, we rely on robust segmentation of lung structures, which is the first and most critical step in the registration pipeline. While small internal datasets exist, they are insufficient to train generalizable deep neural networks.
The NLST dataset offers a unique opportunity due to its large size, diversity of imaging conditions, and focus on low-dose chest CT. We plan to use only the CT image data, without any access to patient identifiers or clinical outcomes. The segmented images will be used to train and evaluate models that can perform with high accuracy and robustness across varied anatomies.
The final models will be integrated into a clinical software system under development, intended for use in interventional pulmonology. The project aligns with the broader goal of improving the precision and safety of minimally invasive procedures using AI and robotics.
Aims
Develop deep learning models for automatic segmentation of lung fields in low-dose chest CT scans.
Train the models using the NLST dataset to ensure generalizability across diverse patient populations.
Validate segmentation performance using quantitative metrics (e.g., Dice coefficient) against internal annotated datasets.
Integrate the segmentation output into our robotic electromagnetic navigation pipeline for thoracic procedures.
Evaluate the downstream impact of segmentation quality on CT-to-patient registration accuracy.
Prepare the system for potential clinical translation as part of a commercial navigation platform.
Collaborators
Dr Stephen Solomon, Memorial Sloan Kettering Cancer Center, New York