Click here to

Session: Early-Career Investigator Symposium [Return to Session]

An Integrated Machine Learning and Biomechanical Modeling Guided Deformable Image Registration Framework for Consistent Parenchymal Tissue Tracking in Lung CT Scans

B Stiehl*, M Lauria, L Naumann, D O'Connell, P Boyle, I Barjaktarevic, D Low, A Santhanam, UCLA, Los Angeles, CA


MO-FG-BRB-10 (Monday, 7/11/2022) 1:45 PM - 3:45 PM [Eastern Time (GMT-4)]

Ballroom B

Purpose: Functional lung avoidance during radiotherapy aims to minimize excess radiation delivered to parenchymal regions with the greatest ventilation. However, conventional deformable image registration techniques do not account for inaccuracies in the parenchyma due to lack of structural information.

Methods: We present an approach that integrates machine learning and biomechanical modeling to predict parenchymal lung deformation during free breathing. The framework takes Fast Helical Free-Breathing Scans (FHFBCT) as input, from which a 5DCT motion model was assembled. Pseudo end-exhalation and end-inhalation CT scans were obtained from the 5DCT model and segmented lung masks were created to identify the lung boundaries and blood vessels. The deformation of these structures was defined by a conventional optical flow registration method and formed the input for the biomechanical model. Using an end-exhalation CT, we then predicted the lung tissue elasticity using a generative adversarial network (GAN) framework. A set of 30 lung datasets consisting of elasticity measured using a well-validated 5DCT model-guided elastography technique were used for training. The GAN-generated elasticity result formed the second input for the biomechanical model. The biomechanical model was then assembled with initial end-exhalation geometry, tissue elasticity, and deformation boundary constraints and employed to perform a deformation of the geometry according to the given elasticity distribution and boundary constraints.

Results: A set of 10 datasets were used for validation. The registration method predicted the lung parenchymal deformation in a biomechanically consistent manner. Agreement between the parenchymal deformation generated by the model-guided elastography approach and the deformation measured from the biomechanical model with given elasticity and boundary constraints was found to be within 1 mm for > 95% of lung voxels on average.

Conclusion: A biomechanically guided deformable image registration provides the precise and consistent measurement of parenchymal tissue deformation that is critical for delineating functional regions within the lung.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by the Tobacco Related Disease Research Program 27IR 0056, NIH R56 1R56HL139767 01A1, Ken and Wendy Ruby Foundation, and the UCLA Department of Radiation Oncology


Registration, Quantitative Imaging, Lung


IM/TH- Image Registration Techniques: Nonrigid biomechanical registration

Contact Email