Deriving radiomics features from CT images using a partial least square correlation technique in the lung cancer setting.
Bagher-Ebadian H, Qixue W, Siddiqui F, Liu C, Ajlouni M, Movsas B, and Chetty IJ. Deriving radiomics features from CT images using a partial least square correlation technique in the lung cancer setting. Med Phys 2017; 44(6):3213.
Purpose: To apply a novel approach, partial least square correlation method, to derive radiomics features from CT images for lung cancer patients. Methods: CT image datasets for 10 patients with lung cancer treated with IMRT (60-Gy in 30-fractions) were investigated. Three-month follow-up CT was used to classify 5 patients without pneumonitis (group-A), and 5 who developed pneumonitis (group-B). Normal lung tissue was evaluated within a 3- cm ring surrounding the PTV, excluding the GTV. 165 radiomics features from the following eight categories were extracted from the 3-cm ring: 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). Partial-Least-Square-Correlation (PLSC) and Singular-Value-Decomposition (SVD) techniques were recruited to characterize relationships between features extracted from three-month follow- up CTs of groups-A and B. PLSC-SVD generated two sets of latent-variable- vectors representing how features are affected. Results: Among 165 features extracted from follow-up CTs of groups-A and B, five feature categories with 20 features showed significant Mean-Absolute-Percent-Change (MAPC): Mean intensity for IBHF (MAPC = 19.00%, p = 0.0229), Energylevel- 3 for DOST (MAPC = 35.38%, p = 0.0356), 4 Energy and 6 Entropy features for 2DWT (MAPC = 25.57%, pFisher < 0.0001), 5 Energy and 1 Entropy feature for 2DGF (MAPC = 53.52%, pFisher < 0.0001), Auto-Correlation and Sum-Average for GLCM (MAPC = 31.92%, pFisher = 0.0039). Follow-up image mean-HU values for patients with pneumonitis were significantly higher (19%) than those without it. Conclusion: This pilot study suggests feasibility of using the PLSC-SVD method for characterization of radiomics features derived from CT-images of lung cancer patients. Results are in support of our hypothesis, that textural features sensitive to image intensity changes resulting from radiation-induced injury are different between patients with and without radiation pneumonitis, whose image HU differences are also different. Ability of certain features to detect changes in image intensity and frequency suggests potential to identify possible response biomarkers for response. A much larger sample size is warranted.