Skip to main content

The use of deep learning feature extraction of CXRs can improve classification of tuberculosis

Friday, September 14, 2018 — Poster Session V

12:00 p.m. – 1:30 p.m.
FAES Terrace


  • O Juarez-Espinosa
  • A Gabrielian
  • E Engle
  • D Hurt
  • A Rosenthal


The TB Portals Program (TBPP, is an international partnership for clinical informatics and advanced research in tuberculosis. The TBPP collects clinical, socioeconomic, genomic and radiological information from tuberculosis patients, especially those with drug-resistant forms of the pathogen. To aid in personalized, fast and precise diagnostic and monitoring of the disease, the automated analysis of X-ray images of chest area (CXRs) is considered very promising. We have analyzed a large collection of CXRs, comparing images from drug-resistant and drug-sensitive forms of tuberculosis. Every image was represented by a vector of values or features. We started with commonly used image descriptors such as texture, pixels, and other general image properties. We expanded the analysis by using deep learning methods to allow for generating additional image features. In this work, two data sets of 600 X-rays were used to classify images. One set was used to test a classifier for X-rays with different resistance to drugs; and the second was used to classify X-rays as normal or abnormal. The features were extracted using two methods. The first method was based on a histogram of gradients, texture data, and local binary pattern; and the second method used a pretrained model for AlexNet. Using the two sets of features, we utilized a machine learning support vector machine (SVM) to classify the images. The results of the classification were compared based on the quality of the classification task. The results of the experiments show that the manual (feature extraction) and AlexNet methods yielded equivalent accuracy. Improving the uniform feature identification and extraction from CXRs promises to aid in patient diagnostics and treatment planning.

Category: Computational Biology