Feature group | Quantity | Description | Rationale |
---|---|---|---|
Laws energy measures | 25 | Response to 5-pixel × 5-pixel filter targeting combination of specific textural enhancement patterns in the x and y directions. Descriptors include all combinations of five 1D filters: level (L), edge (E), spot (S), wave (W), and ripple (R). | May possibly detect patterns of heterogeneous enhancement and abnormal structure; have previously been shown to enable quantification of TILs by lung CT [57]. |
Gabor features | 48 | Detection of edges through response to Gabor wavelet features. Each descriptor quantifies response to a given Gabor filter at a specific frequency (f = 0, 2, 4, 8, 16, or 32) and orientation (θ = 0 degrees, 22.5 degrees, 45 degrees, 67.5 degrees, 90 degrees, 112.5 degrees, 135 degrees, 167.5 degrees). | May possibly capture changes in tumor microarchitecture on account of glandular morphology or detect the presence of TILs [57]. TILs have been shown to be prognostic of better survival and NAC response [54]. |
Haralick features | 13 | Quantify heterogeneity and entropy of local intensity texture as represented by the gray-level co-occurrence matrix within a 5-pixel × 5-pixel window. | Regional changes in Haralick features following treatment have been shown to predict pCR in breast cancer [19]. |
Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe) features | 13 | Apply Haralick metrics to dominant intensity gradient orientations within a 5-pixel × 5-pixel window, quantifying patterns of local gradient alignment [59, 60]. Some descriptors quantify homogeneity of gradient orientations (e.g., CoLlAGe information measure of correlation), whereas others compute their disorder (e.g., CoLlAGe entropy). | CoLlAGe entropy has previously been demonstrated to be effective in distinguishing breast cancer subtypes [59, 60]. |