Room: AAPM ePoster Library
Purpose: To develop a method for predicting lung elasticity distribution using only a single low dose CT scan, enabling elasticity distributions to be widely employed in diagnostic and research applications.
Methods: We developed a machine learning based approach that predicts a 3D lung tissue elasticity distribution given only a low dose breath-hold CT scan. 10 patient low-dose free breathing CT (FBCT) datasets were used to simulate ultra low-dose equivalents, which were in turn used to generate low-dose 5DCT motion models. These motion models were then used to generate end-inhalation and end-exhalation CT scans, from which tissue elasticity was estimated with a biomechanically guided inverse elasticity approach. A machine learning process (constrained Generalized Adversarial Neural Network (cGAN)) learned the lung tissue elasticity in a supervised manner for the simulated low dose end-expiration CT. The estimated elasticity was then validated using (a) direct comparison with the 5DCT-based elasticity data via an L2-norm, and (b) regenerating synthetic 4DCTs and comparing with the ground-truth 4DCTs using 3 image similarity metrics: Mutual Information (MI), Structured Similarity Index (SSIM), and Normalized Cross Correlation (NCC).
Results: For the training data set consisting of 8 low dose CT scans, we obtained a learning accuracy of 0.6 kPa. For the validation dataset consisting of 2 scans, we obtained an accuracy of 0.9 kPa.
Conclusion: The cGAN generated lung tissue elasticity can be effectively estimated given a low dose end-expiration CT image with clinically acceptable accuracies. This framework opens research avenues, where CT-based elastography can be used in a diagnostic setup for characterizing and phenotypic lung diseases.