Author: Huay Din, Grace Jianan Gang, Grace Hyun Kim, Michael F. McNitt-Gray, Joseph W. Stayman, Yijie Yuan 👨🔬
Affiliation: Johns Hopkins University, John Hopkins University, University of Pennsylvania, David Geffen School of Medicine at UCLA 🌍
Purpose:
Radiomics rely on quantitative features to discern underlying biological signatures. However, feature dependence on the imaging systems themselves hampers the creation of reproducible and generalizable models. We propose a novel framework to remove the effects of system blur and image noise on radiomic calculations. Prior work has validated the approach in CT for histogram- and gray-level co-occurrence matrix (GLCM) radiomics. We extend the framework to run length matrices (RLM), wavelet, and Fourier-based radiomics.
Methods:
Under a locally linear and shift-invariant systems, the standardization procedure involves two deconvolutions: an image-domain deconvolution to remove the effect of blur and a radiomics-domain deconvolution to remove the effect of noise. The approach can be directly applied to radiomics from linearly transformed images (wavelet, Fourier). To recover RLM radiomics, we recognize that the i-th column of the RLM is the diagonal of the i-dimensional joint probability distribution of neighboring pixels; hence, each column can be recovered by an i-dimensional radiomics-domain deconvolution. For validation, both micro-CT scans of a 3D-printed texture phantom (GLRLM, wavelet) and simulated power law noise were used as ground truth. Five different levels of blur, noise-correlation/magnitude were simulated with 100 noise realizations each.
Results:
Across most imaging conditions, both the recovered radiomics and features are closer in shape and values to their ground truth counterparts: The blurred and noisy vs. recovered average MSE and standard deviation for the RLMs are 2.9e4±1.3e4 vs. 1.2e5±2.3e3; Haar-HL wavelets 1.7e2±7.4e1 vs. 2.5e7±1.8e7; log-difference magnitude Fourier histograms 6.5e-2±3.6e-2 vs. 1.2e-1±1e-2; and the average absolute percentage difference across all classes of radiomics is 2.1e2±2e3 vs. 3.4e3±6.1e3 respectively.
Conclusion:
The work establishes an analytical standardization approach capable of recovering ground truth radiomics values given known system blur and image noise distributions. Incorporating the approach prior to radiomics model building can eliminate system-dependence and improve model robustness and generalizability.