Propagation of Low-Dose-Induced Noise Characteristics Across Deep Learning Layers in CT Imaging.
Authors
Affiliations (9)
Affiliations (9)
- Department of Medical Imaging and Radiation Therapy, Hanze University of Applied Sciences, Groningen, 9714 CA, The Netherlands. [email protected].
- Department of Radiotherapy, University of Groningen, University Medical Centre Groningen, 9713 GZ, Groningen, The Netherlands. [email protected].
- Research Group Healthy Ageing, Allied Health Care and Nursing, Hanze University of Applied Sciences, Groningen, 9714 CA, The Netherlands. [email protected].
- Research Group Healthy Ageing, Allied Health Care and Nursing, Hanze University of Applied Sciences, Groningen, 9714 CA, The Netherlands.
- Faculty of Science and Engineering, Undergraduate School of Science and Engineering, University of Groningen, 9747 AG, Groningen, The Netherlands.
- Department of Medical Imaging and Radiation Therapy, Hanze University of Applied Sciences, Groningen, 9714 CA, The Netherlands.
- Department of Radiotherapy, University of Groningen, University Medical Centre Groningen, 9713 GZ, Groningen, The Netherlands.
- Department of Radiology, University of Groningen, University Medical Centre Groningen, 9700 RB, Groningen, The Netherlands.
- Data Science Center in Health, University Medical Centre Groningen, 9713 GZ, Groningen, The Netherlands.
Abstract
Dose reduction in computed tomography (CT) imaging impacts diagnostic accuracy for human observers and potentially artificial intelligence (AI) systems. As noise propagates through an AI model, its characteristics change, possibly influencing model performance. This study aims to develop a configurable low-dose CT simulation tool in image-space and analyze noise behavior in deep learning (DL) layers. The noise characteristics were quantified using the noise power spectrum (NPS) within homogeneous phantom regions. NPS area under the curve (intensity) and median frequency (texture) were reported. Low-dose CT images were simulated using a custom program that was calibrated for configurable noise intensity and texture. Noise was generated at texture levels. Pre-trained U-Net features were extracted from reference and simulated low-dose images, enabling NPS-based analysis of noise propagation across deep learning layers. NPS from simulated low-dose CT images matched reference NPS for various intensity and texture values. Feature analysis revealed progressive convergence of noise characteristics across deep learning layers, with batch normalization mitigating amplification observed after convolution and ReLU activation. Differences in noise texture between low-dose simulations converge along the U-Net contracting path. The image-space simulation tool offers a viable alternative for researchers lacking access to raw data, supporting future AI-focused dose optimization studies. However, whether these outcomes translate to segmentation performance remains unclear and warrants further investigation.