Click here to

Session: Deep Learning Response Prediction, Diagnosis, and Modeling [Return to Session]

A Generative Adversarial Transfer Learning Framework for Predicting Lung Diaphragm and Rib-Cage Mechanics From Free and Forced Breathing Lung Imaging

A Santhanam1*, B Stiehl2, M Lauria3, L Naumann4, I Barjaktarevic5, M McNitt-Gray6, D Low7, (1) University of California, Los Angeles, Los Angeles, CA, (2) ,Los Angeles, CA, (3) UCLA, Los Angeles, CA, (4) ,Los Angeles, CA, (5) University of California, Los Angeles, Los Angeles, ,(6) David Geffen School of Medicine at UCLA, Los Angeles, CA, (7) UCLA, Los Angeles, CA


WE-C930-IePD-F6-5 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Lung deformation is guided by the diaphragm and rib-cage motion mechanics(DRM) during free and forced-breathing. In this abstract, we present a novel framework to predict the DRM using a single CT.

Methods: A set of 30 lung cancer subjects with free breathing images and 12 patients with forced breathing images were considered for this study. For each lung cancer patient, 25 FHFBCTs were acquired and a 5DCT model was assembled. For patients with forced breathing, a pair of lung images at Residual Volume (RV) and Total Lung Capacity (TLC) were acquired with the former being considered as the reference CT. The DRMs were encoded by characterizing the voxels adjacent to the lung geometry. GAN transfer learning: First, a modified constrained Generative Adversarial Learning framework was employed to predict the lung DRM for free breathing. The adversarial loss included a weight distribution for the parenchymal, vascular, and the DRM. The adversarial loss enabled predicting the lung DRM with <5% error. A novel transfer learning was then added to the trained models. For both the discriminator and generator DNNs, two additional CNN layers were added to modeling the forced breathing maneuvers. These two layers were trained using the forced breathing scans. The drop-out rates were set to facilitate an accurate prediction of the forced-breathing DRM.

Results: The learning accuracy was observed to have a <5% mean error for both the free and forced breathing lung images. The DRM results were then incorporated into a biomechanical model to predict the lung deformation, which yielded a 87% voxels to have <1 mm error in simulating the lung breathing during free and forced breathing

Conclusion: A conditional GAN enables an optimal lung CT based DRM prediction, which demonstrates the ability to deform breath-hold and RV CT scans for both free and forced breathing.

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


Lung, Image-guided Therapy, Respiration


Not Applicable / None Entered.

Contact Email