Classification of DPLDs in HRCT scans: a comparative study of texture analysis methods and a novel statistical and graph-based approach

Abstract

Problems of texture classification are consistently challenging once the patterns of different instances can be very similar. Moreover, the descriptors need to be invariant to rotations, scale, and lighting variations. In the context of medical imaging, this group of methods can aid in diagnosing patients as part of the concept of Computer-Aided Diagnosis (CAD). In this paper, we review methods for texture classification in the context of classifying Diffuse Parenchymal Lung Diseases (DPLDs) on High-Resolution Computed Tomographies (HRCTs) and propose a new method that uses concepts of complex networks and statistical metrics. Our approach is based on mapping the input image into multiscale graphs and extracting the closeness centrality metric. We transform the multiscale closeness centrality images into one matrix that encapsulates local and global texture information. From the matrix, we extract a feature vector that represents a DPLD pattern. This vector is then combined with Haralick and Local Binary Pattern descriptors to generate the final feature vector. Once this process characterizes all the images, we go through a classification step to recognize the image. We analyze the proposed approach’s performance by comparing it with other texture analysis methods and discussing its metrics for each class (DPLD pattern) of the dataset. After the evaluation of different methods and our proposed method, it is possible to conclude the effectiveness of our approach to aid the diagnose process of DPDLs. Furthermore, we can highlight our technique as an aid on the problem of diagnosing patients with COVID-19.