Click here to

Session: Machine Intelligence in Image Processing and Motion Correction I [Return to Session]

A Systematic Comparison of DNN-Based Lung CT Elastography Using Auto-Encoder, U-Net, Conditional and Cycle Generative Adversarial Network Techniques

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


TH-E-BRC-1 (Thursday, 7/14/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Function preserving lung radiotherapy efforts can benefit from a quantitative characterization of the lung tissue’s effective elasticity generated from breath-hold lung CTs.

Methods: We conducted a systematic comparison of DNN-based lung CT elastography using four state-of-art techniques, including auto-encoder, U-Net, and conditional and cycle generative adversarial networks (GANs). A set of 30 lung cancer subjects were considered for this study. For each patient, 25 FHFBCTs were acquired and a 5DCT motion model was generated according to 5DCT protocol. A high-resolution lung biomechanical model was then assembled using a reference end-exhalation CT (generated from 5DCT model) and deformation vector field (DVF) representing end-exhalation to end-inhalation lung deformation. Using the biomechanical model implementation, parenchymal tissue elasticity was measured according to a well-validated elasticity estimation workflow. The end-exhalation CT images and associated elasticity formed the training data for the DNN-based elastography. The auto-encoder consisted of a 14 layer fully connected DNN network which predicted the lung elasticity from the reference CT. The U-net framework consisted of a similarly structured DNN along with connections between the lateral layers. Both the conditional and CycleGANs consisted of discriminator and generator networks, where the generator network predicted the elasticity while the discriminator facilitated adversarial learning. The CycleGAN also predicted the reference CT from the lung elasticity thereby ensuring inverse consistency.

Results: Both auto-encoder and U-net frameworks were able to predict lung elasticity at ~8% error. The adversarial learning approaches, conditional and Cycle, led to observed prediction errors of < 3% and < 5%, respectively. The inverse consistency constraint in the cycle GANs yielded a lower accuracy because of a low-rank correlation between the elasticity and the reference CT.

Conclusion: A conditional GAN enables optimal CT-based lung elastography prediction from a single breath-hold CT, which can inform function preserving lung radiotherapy efforts using conventional lung imaging.

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


Quantitative Imaging, Lung, Image Analysis


IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

Contact Email