Click here to

Session: AI in CT and CBCT: Image Enhancement and Synthesis [Return to Session]

Evaluation of Image-Domain Training Frameworks for Deep Learning Based Denoising and Deconvolution

P VanMeter1*, S Hsieh1, J Marsh1, N Huber1, A Ferrero1, C McCollough1, (1) Department of Radiology, Mayo Clinic, Rochester, MN


TU-D-TRACK 3-7 (Tuesday, 7/27/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Purpose: Spatially matched full-dose/low-dose (FD/LD) pairs are often used as training input and target for supervised learning. Training data is commonly simulated using projection noise insertion; however, inability of many institutions to access CT projection data motivates development of widely accessible image domain methods. In this work, we evaluated two image domain training frameworks for CNN-based dual-energy denoising and deconvolution applied to ex-vivo kidney stone specimens.

Methods: We compared two image domain training frameworks, a Generalizable Noise and Artifact Reduction Network (GARNET) and an isotropic colored noise insertion technique (Iso-NI), with an existing reference technique using projection-domain noise insertion (PD-NI). GARNET training data was synthesized using the following procedure: noise-only images were extracted from phantom scans, re-scaled to the desired dose, and added back to FD images to mimic LD images for training. Iso-NI training data was synthesized by backprojecting filtered Gaussian random noise into an image, which was then blurred to match the noise power spectrum of the LD images. Patches derived from FD/LD pairs were used to train a separate CNN for each technique, and denoised/deconvolved kidney stone images were inferred. Inferred images were input into a dual-energy stone composition algorithm to evaluate the performance in a clinical task.

Results: GARNET and the PD-NI reference technique were found to have similar noise power spectra, whereas Iso-NI inferred images were slightly noisier. All methods resulted in comparable, excellent performance for kidney stone composition quantification (9 out of 11 mixed stones properly identified by all methods vs 0/11 in the original data).

Conclusion: Inferred images using both GARNET and Iso-NI trained CNNs were shown to significantly reduce noise and improve stone quantification relative to the original images. These methods provide a suitable substitute to projection domain noise insertion when access to proprietary projection data is unavailable.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH grant R01 EB028591.



    Noise Reduction, Quantitative Imaging, CT


    IM- CT: Quantitative imaging/analysis

    Contact Email