Related Experiment Video
Updated: Jun 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Unsupervised Denoising in Spectral CT: Multi-Dimensional U-Net for Energy Channel Regularisation
Raziye Kubra Kumrular1, Thomas Blumensath1
1Institute of Sound and Vibration Research, Department of Engineering and the Environment, University of Southampton, University Rd., Southampton SO17 1BJ, UK.
Sensors (Basel, Switzerland)
|October 26, 2024
Summary
Noise2Inverse image denoising effectively reduces noise in spectral computed tomography (CT) data. This unsupervised deep learning method enhances quantitative material identification without complex parameter tuning.
Area of Science:
- Medical Imaging
- Computational Imaging
- Image Processing
Background:
- Spectral Computed Tomography (CT) offers energy-dependent X-ray attenuation data using Photon Counting Detector (PCD) technology.
- Increased noise in spectral CT, due to lower photon counts across multiple energy channels, hinders quantitative material identification.
- Effective noise reduction is crucial for advancing spectral CT applications in industry, medicine, and research.
Purpose of the Study:
- To investigate the efficacy of the Noise2Inverse image denoising approach for noise reduction in spectral CT.
- To develop and evaluate an unsupervised deep learning model for spectral CT noise removal.
- To assess the performance of the proposed method against existing techniques for quantitative material identification.
Main Methods:
- Implemented an unsupervised deep learning model based on a multi-dimensional U-Net architecture.
- Utilized a block-based training approach with modifications for energy-channel regularization.
- Conducted experiments on simulated spectral CT phantoms and a real biological sample with a K-edge.
Main Results:
- The Noise2Inverse approach demonstrated superior performance in noise reduction compared to unsupervised Low2High and total variation-constrained iterative reconstruction methods.
- Quantitative metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Contrast-to-Noise Ratio (CNR) confirmed the effectiveness of the denoising method.
- The model achieved significant noise reduction without requiring intricate parameter tuning.
Conclusions:
- The Noise2Inverse denoising approach is a highly effective and user-friendly solution for noise reduction in spectral CT.
- This method significantly improves quantitative material identification capabilities in spectral CT imaging.
- The unsupervised deep learning strategy offers a promising direction for advancing spectral CT data analysis and applications.
