Purpose: This study is to investigate the potential of deep learning (DL) in CT dose reduction while preserving image texture when compared to iterative reconstruction (IR) alone.
Methods: The upper abdomen of a live 4-year-old sheep was scanned under anesthesia on a Siemens Force CT scanner at different exposures (120 kV, mA from 25 to 250). Images were reconstructed using both Filtered Back Projection (FBP) and Advanced Modeled Iterative Reconstruction (ADMIRE) at each exposure with various strengths (1-5). A recently-developed modularized deep neural network (MAP-NN) with 5 modules was used for CT imaging at each exposure level via progressive denoising. Region of Interest (ROI) were manually delineated over the liver for all image sets. Radiomic features were retrospectively extracted from the ROI with IBEX. The concordance correlation coefficient (CCC) was applied to quantify the agreement between any two sets of radiomic features. The structural similarity index measure (SSIM) was used to measure the similarity between any two images.
Results: A total of 113 features are extracted. For FBP, the CCC shows consistent decrease with lower exposure. Both IR and DL have the potential to increase CCC of radiomic features with desired features (e.g., features at standard exposure with FBP), depending on DL/IR strength. DL has higher potential than IR to improve radiomic features at extremely lower exposure level. The relationship between exposure and IR/DL strength for dose reduction while preserving image quality is nonlinear. SSIM decreases with progressive increases of DL or IR strength. DL has higher potential than IR to preserve image structure at higher strength, independent of exposure.
Conclusion: Using a multi-exposure experiment in a live animal model, deep learning shows better potential than iterative reconstruction to further reduce radiation dose while preserving structural similarity and feature correlation.
CT, Radiation Dosimetry, Image Analysis