Texture analysis (GLCM)

Texture analysis is frequently used to classify radiological images²¹. Wu et al.²² used texture features for classifying fibrosis stage and necroinflammatory activity in the liver. Generally, texture features from statistical approaches include histogram, gradient, gray-level co-occurrence matrix (GLCM), etc. Considering the spatial correlations between voxels, the GLCM, which describes pairwise arrangement of voxels with the same gray-level, was used in this study to extract information of local similarities.

Co-occurrences of pairs of voxels are defined using relative distance²¹. In addition, the grayscale value of each voxel is quantized to Ng gray levels. Therefore, a matrix of relative frequencies consists of Pk,l, the probability of two neighboring voxels at a distance d and an angle α, having the intensity scales k, l (k,l=1,2,...,Ng), respectively.

Haralick et al.²¹ proposed fourteen texture features extracted from the GLCM for quantitative analysis of image texture. P. Mohanaiah et al.²³ showed that four second order features provide high discrimination accuracy in image analysis: Angular Second Moment (energy), Correlation, Entropy, and the Inverse Difference Moment (IDM). They are defined as:

where μk=Σk,lNgkPk,l, μl=Σk,lNglPk,l and σk=Σk,lNgPk,l(k-μk)2, σl=Σk,lNgPk,l(l-μl)2. These four features were summarized as texture features for classification and extracted using the python package Radiomics²⁴.