Physics-Informed Machine Learning for Estimating Pulmonary Perfusion From Non-Contract 4DCT
Y Liu1*, A Nowacki1, R Castillo2, Y Vinogradskiy3, G Nair4, C Stevens4, E Castillo1, (1) University of Texas at Austin, Austin, TX, (2) Emory University, Atlanta, GA, (3) Thomas Jefferson University, Philadelphia, PA, (4) William Beaumont Hospital, Royal Oak, MI
Presentations
SU-E-201-7 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room 201
Purpose: Novel methods have been developed for functional avoidance that proposes to use 4DCT to derive lung ventilation and perfusion images. Previous methods for quantifying lung perfusion on non-contrast 4DCT rely either on HU-based physical models or black-box deep learning models. While deep learning typically achieves higher accuracies in image processing tasks, physical models provide a rationale for model predictions. The purpose of this study is to introduce a biophysics-informed machine learning method for estimating pulmonary perfusion from non-contrast 4DCT. Our approach is designed to combine the predictive power of neural networks with the interpretability of physical modeling, with the goal of providing high-fidelity information for radiotherapy functional avoidance planning.
Methods: Simulation 4DCT scans and SPECT-Perfusion scans for 42 non-small cell lung cancer patients were used to train and validate a 3-layer, fully connected, artificial neural network (ANN). Similar to existing physics-based models for CT-perfusion, the ANN takes as inputs the spatially corresponding inhale/exhale lung densities and Jacobian measured volume changes for the five lung lobes. The output layer predicts the percent of total lung perfusion within each lobe. SPECT-Perfusion images were used for ground truth. The ANN was trained using the stochastic gradient descent optimizer with grid search. Leave-one-out cross-validation was applied to estimate prediction quality.
Results: The average mean square errors resulting from the leave-one-out process was 5.71±2.77%. The median(interquartile range) of the Spearman correlations between ground truth and predictions was 0.7(0.5).
Conclusion: Our proposed physics-informed ANN generates spatial correlations with SPECT-Perfusion that provide improved correlation compared to existing methods. Moreover, the approach is based on lung density and volume change measurements which are well-known physical quantities that have been shown to correlate with disease and functional defects. Therefore, the developed ANN represents a novel and interpretable machine learning methodology for quantifying perfusion from non-contrast CT.
Keywords
Taxonomy
Contact Email