On the impact of smoothing and noise on robustness of CT and CBCT radiomics features for patients with head and neck cancers.

PURPOSE: We investigated the characteristics of radiomics features extracted from planning CT (pCT) and cone beam CT (CBCT) image datasets acquired for 18 oropharyngeal cancer patients treated with fractionated radiation therapy. Images were subjected to smoothing, sharpening, and noise to evaluate changes in features relative to baseline datasets.
METHODS: Textural features were extracted from tumor volumes, contoured on pCT and CBCT images, according to the following eight different classes: intensity based histogram features (IBHF), gray level run length (GLRL), law's textural information (LAWS), discrete orthonormal stockwell transform (DOST), local binary pattern (LBP), two-dimensional wavelet transform (2DWT), Two dimensional Gabor filter (2DGF), and gray level co-occurrence matrix (GLCM). A total of 165 radiomics features were extracted. Images were post-processed prior to feature extraction using a Gaussian noise model with different signal-to-noise-ratios (SNR = 5, 10, 15, 20, 25, 35, 50, 75, 100, and 150). Gaussian filters with different cut off frequencies (varied discreetly from 0.0458 to 0.7321 cycles-mm-1 ) were applied to image datasets. Effect of noise and smoothing on each extracted feature was quantified using mean absolute percent change (MAPC) between the respective values on post-processed and baseline images. The Fisher method for combining Welch P-values was used for tests of significance. Three comparisons were investigated: (a) Baseline pCT versus modified pCT (with given filter applied); (b) Baseline CBCT versus modified CBCT, and (c) Baseline and modified pCT versus baseline and modified CBCT.
RESULTS: Features extracted from CT and CBCT image datasets were robust to low-pass filtering (MAPC = 17.5%, pvalFisher¯ = 0.93 for CBCT and MAPC = 7.5%, pvalFisher¯ = 0.98 for pCT) and noise (MAPC = 27.1%, pvalFisher¯ =  0.89 for CBCT, and MAPC = 34.6%, pvalFisher¯ = 0.61 for pCT). Extracted features were significantly impacted (MAPC=187.7%, pvalFisher¯ < 0.0001 for CBCT, and MAPC = 180.6%, pvalFisher¯ < 0.01 for pCT) by LOG which is classified as a high-pass filter. Features most impacted by low pass filtering were LAWS (MAPC = 11.2%, pvalFisher¯ = 0.44), GLRL (MAPC = 9.7%, pvalFisher¯ = 0.70) and IBHF (MAPC = 21.7%, pvalFisher¯ = 0.83), for the pCT datasets, and LAWS (MAPC = 20.2%, pvalFisher¯ = 0.24), GLRL (MAPC = 14.5%, pvalFisher¯ = 0.44), and 2DGF (MAPC=16.3%, pvalFisher¯ = 0.52), for CBCT image datasets. For pCT datasets, features most impacted by noise were GLRL (MAPC = 29.7%, pvalFisher¯ = 0.06), LAWS (MAPC = 96.6%, pvalFisher¯ = 0.42), and GLCM (MAPC = 36.2%, pvalFisher¯ = 0.48), while the LBPF (MAPC = 5.2%, pvalFisher¯ = 0.99) was found to be relatively insensitive to noise. For CBCT datasets, GLRL (MAPC = 8.9%, pvalFisher¯ = 0.80) and LAWS (MAPC = 89.3%, pvalFisher¯ = 0.81) features were impacted by noise, while the LBPF (MAPC = 2.2%, pvalFisher¯ = 0.99) and DOST (MAPC = 13.7%, pvalFisher¯ = 0.98) features were noise insensitive. Apart from 15 features, no significant differences were observed for the remaining 150 textural features extracted from baseline pCT and CBCT image datasets (MAPC = 90.1%, pvalFisher¯ = 0.26).
CONCLUSIONS: Radiomics features extracted from planning CT and daily CBCT image datasets for head/neck cancer patients were robust to low-power Gaussian noise and low-pass filtering, but were impacted by high-pass filtering. Textural features extracted from CBCT and pCT image datasets were similar, suggesting interchangeability of pCT and CBCT for investigating radiomics features as possible biomarkers for outcome.