Blockwise Classification of Lung Patterns in Unsegmented CT Images

Abstract

Diagnosis of lung diseases is usually accomplished by detecting abnormal characteristics in Computed Tomography (CT) scans. We report an initial study for classifying texture patterns in High-Resolution lung CTs using the Completed Local Binary Pattern (CLBP) descriptor with a Support Vector Machine (SVM). The main contribution of the proposed method is that it does not depend on a previously segmented lung, as it performs a coarse segmentation by classifying body areas outside the lungs. The classified patterns are: non lung, normal lung tissue, emphysema, ground-glass opacity, fibrosis and micronodules. Using image blocks of 32 × 32 pixels, extracted from a public dataset with 113 patients, correct blockwise classification of non lung patterns was achieved with an accuracy of 98.91%. Regarding normal and pathological lung patterns, a mean accuracy of 91.81% was obtained. This is similar to the reported results in literature which used a pre-segmented lung.

Holistic Interstitial Lung Disease Detection using Deep Convolutional Neural Networks: Multi-label Learning and Unordered Pooling

“…Detecting ILD patterns in CT imaging is commonly treated as a texture recognition and classification problem in many previous studies [22,8,23,24,25]. Moreover, texture based visual representation is adopted inside local image regions of interest (ROIs) or volumes of interest (VOIs) via extracting rectangular image patches, when a 2D and 3D CT imaging modality is used, respectively.…”

“…There are many types of hand-crafted image features that are adopted for ILD classification, such as filter banks [22,8,23], local binary patterns (LBPs) [24,25], morphological operators followed by geometric measures, histogram of oriented gradients [8], texton based approaches [32], and wavelet and contourlet transforms [33,34]. 2D texture features have also been extended into three dimensions [35,36,28].…”

“…2D texture features have also been extended into three dimensions [35,36,28]. Typical feature encoding scheme and classifiers include bag of words [37], support vector machines (SVMs) [24,35,34], random forest [23] and k-nearest neighbors (kNN) [22].…”