Author: Magdalena Bazalova-Carter, James Day, Xinchen Deng 👨🔬
Affiliation: University of Victoria 🌍
Purpose:
Ring artifacts in Photon-Counting Computed Tomography (PCCT) images can degrade image quality. this study aims to suppress ring artifacts with a novel autoencoder-based framework that leverages nonlinear dimensionality reduction.
Methods:
We use a deep autoencoder to project noisy sinograms into a low-dimensional latent space, filtering out artifact-prone features. The autoencoder is applied to capture structures that represent the primary signal. This reliably solves artifact removal through dimension reduction. The model consists of an encoder that compresses the input data and a decoder that reconstructs the artifact-reduced sinograms. The model is applied to noisy sinograms in an unsupervised manner, minimizing a combined loss function between the input and the output. The model is applied on 120kV PCCT images of a phantom and biological samples acquired on a benchtop system at multiple energy bins with energy thresholds of 24, 34, 44, 54, 64, 74, 84keV. Image entropy and line profile are used as quantifications of ring artifacts.
Results:
The deep autoencoder effectively reduced ring artifacts while preserving structural details in PCCT images. Visual inspection of PCCT images and their derivatives of sinograms of pig trotter and ox tail PCCT images demonstrated notable reductions of ring artifacts. The approach is effective across datasets and energy bins. Quantitative evaluation supported these findings in a calibration phantom with uniform regions: Image entropy values were reduced from 3.73 to 3.55 and 3.71 to 3.56 for slices with prominent ring artifacts. Line profile analysis of corrected images also revealed smoothed intensity fluctuations. The standard deviation of the lines reduced from 504.57 to 496.72 and 507.73 to 502.81, respectively.
Conclusion:
The preliminary results show that the deep autoencoder approach demonstrates robust capability for ring artifact suppression in PCCT images by leveraging nonlinear dimensionality reduction. This approach preserves critical structural details and improves image uniformity.