LT-LBP-Based Spatial Texture Feature Extraction with Deep Learning for X-Ray Images

Abstract

The novel logarithmic transformation of Local Binary Pattern (LT-LBP) for texture feature extraction is used with Deep Learning (DL) models. The Computational complexity in the DL models asymptotically increases with the data sesssst and the depth of the layers in Convolutional Neural Networks (CNN). Pre-processing the images before training the DL model holds a strong impact on the operation speed and throughput. The proposed work encompasses LT-LBP for spatial texture feature extraction and is combined with the DL models. LT-LBP reduces the computation while extracting the texture information from the medical images. The computational cost for the model has been reduced in this study. The experiments were done with various DL models such as inception V3, Dense Net, Mobile Net, Efficient Net (B0-B7) VGG16, and RESNET50, and the results were analyzed. The features were extracted using LT-LBP and given as input to the DL models for classification. The retrospective study was done with 118 COVID-19 X-ray images collected from Chengalpattu Medical College Hospital, Chennai, and the remaining 8,224 images were taken from Kaggle. The combination of LT-LBP with RESNET50 shows better performance when compared to other models with an accuracy of 87%. The Area Under the Curve (AUC) for the proposed model is 88, 83 and 86% for COVID, non-COVID, and pneumonia classifications.