A Comparative Study of Texture Analysis Methods on the Classification Problem of DPLDs in HRCT scans

Texture classification is challenging because patterns from different instances can be very similar. Furthermore, the descriptors need to be invariant to rotations, scale, and lighting variations. In this work, we analyzed three different class of methods for texture features extraction: i. Descriptive statistics; ii. Spectral analysis; and iii. Networks that use a layered architecture to characterize images. Various methods of each class were executed to extract features to describe the texture of diffuse parenchymal lung diseases (DPLDs) in computed tomography (CT) scans. Performance analysis was conducted for each class in the dataset. After analyzing the results, it was observed that the T-CNN, the PCANet and the Haralick Attributes presented the best performance.