Comparing Normalizing Flow and Generative Adversarial Networks to Mitigate Effects of Variations in CT Acquisition and Reconstruction
L Wei1, A Yadav2*, W Hsu2, (1) Department of Electrical & Computer Engineering, University of California, Los Angeles, CA (2) Medical & Imaging Informatics, Department of Radiological Sciences, David Geffen School of Medicine at UCLA, Los Angeles, CA
Presentations
TH-D-207-2 (Thursday, 7/14/2022) 11:00 AM - 12:00 PM [Eastern Time (GMT-4)]
Room 207
Purpose: While quantitative image features (QIFs) have been shown to predict disease severity and progression, CT acquisition and reconstruction variations may result in QIFs with poor reproducibility. Various image synthesis methods have been proposed to harmonize image data, mitigating the effect of these variations. We assess the impact of harmonizing image data using generative adversarial networks (GANs) and a normalizing flow alternative approach.
Methods: Commonly used state-of-the-art image synthesis methods such as GANs learn to generate a single output that mimics a target image (e.g., a scan acquired at 100% dose, reconstructed using a medium kernel) given an input image (e.g., a scan acquired at 25% dose, reconstructed using a smooth kernel). However, image synthesis problems, going from a lower resolution to a higher resolution image, are ill-posed and have multiple possible solutions. A normalizing flow-based approach learns the explicit conditional density and can output the entire spectrum of outputs, reflecting the uncertainty of the problem. We evaluate the performance of GAN-based and flow-based models by 1) comparing image quality metrics on a denoising task using the AAPM-Mayo Clinical Low-Dose CT Grand Challenge dataset and 2) assessing the consistency in nodule detection performance across 86 low-dose CT chest scans acquired at our institution and reconstructed using different kernels.
Results: The flow-based model achieved 6% better perceptual quality than GANs while achieving a similar peak signal-to-noise ratio and structural similarity. Qualitatively, GAN-based methods yielded streaking artifacts not observed in the flow-based method. Consistency, assessed using concordance correlation coefficient, for the flow-based method were higher than GANs (0.993±0.003 vs. 0.879±0.044) when comparing the performance of a nodule detection algorithm on a harmonized image versus the reference.
Conclusion: Normalizing flow is a promising alternative to GANs in mitigating the effect of differences in CT acquisition and reconstruction.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- Informatics: Informatics in Imaging (general)
Contact Email