Investigating Physics-Based Standardization of Radiomics from Normal and Fibrotic Lung Regions across CT Reconstruction Kernels and Radiation Doses
Abstract
Purpose
To investigate whether physics-based radiomics standardization can increase agreement in radiomics extracted from radiologist-annotated regions of diffuse lung disease (DLD) phenotypes across CT reconstruction kernels and radiation doses.
Methods
Raw projection data of 3 DLD patients receiving a thoracic CT exam was reconstructed with 5 kernels (smooth, medium-smooth, medium, medium-sharp, and sharp) and 4 radiation dose levels relative to the original (100%, 75%, 50%, and 25%). Normal and fibrotic regions were annotated by a thoracic radiologist on the 100% dose-medium reconstruction, which was used as reference for this investigation. A physics-based radiomics standardization approach that deconvolves blur of a non-reference kernel from an image and reconvolves it with the reference kernel blur, then deconvolves noise from the histogram and GLCM of the resulting image was used to standardize radiomics to those from the reference reconstruction. Within annotations, 12x12 patches were sampled, and radiomic features were measured before and after standardization. Feature values before and after standardization were assessed for agreement with reference using the ratio of number of patches with feature values within 5%, 10% and 20% of the reference value, and concordance correlation coefficient (CCC).
Results
There were 38,121 normal and 62,201 fibrotic 12x12 patches generated. Before standardization, no condition achieved >50% of features within 5% of reference; after standardization, 50% (10/20) and 45% (9/20) of conditions for normal and fibrotic regions exceeded this threshold, respectively. Percent of features with CCC > 0.9 ranged from 7-93% across reconstruction conditions. Standardization yielded higher CCC (58-93%) than before standardization (7-62%). Normal tissue demonstrated higher consistent agreement than fibrotic tissue after standardization. Higher dose levels and kernels closer to reference (Br40d, Br59d) achieved better agreement.
Conclusion
Physics-based standardization of radiomics can increase agreement with radiomics extracted from normal and fibrotic lung regions across a range of CT reconstruction kernels and doses.