Skip to main content
 

Comparative study for classification of radiology images for Tuberculosis patients

Friday, September 16, 2016 — Poster Session IV

12:00 p.m. – 1:30 p.m.
FAES Terrace
NIAID
COMPBIO-15

Authors

  • A Gabrielian
  • E Engle
  • O Juarez-Espinosa

Abstract

Radiology imaging could offer much needed timely information about the disease and the patient, useful for diagnosis and personalized treatment. Automatic analysis of chest X-Ray and CT images allows to search for features correlating with various aspects of Tuberculosis (TB) as a disease and as a continuous process of interaction of pathogen and a host. We are researching ways to improve automatic classification of chest X-Ray images for TB patients. This study compares ability of several machine learning algorithms (SVM, Decision Trees, Clustering, and finally Deep Learning) to distinguish between patient groups with different resistance to common TB drugs. Six hundred images were assigned to five patient categories based on drug resistance: 1) Extreme Drug Resistance; 2) Multidrug Resistance; 3) Sensitive; 4) Mono Drug Resistance; and 5) Poly Drug Resistance. The images and the corresponding annotations were collected as part of on-going international collaboration, aimed at providing comprehensive IT solutions for health care, researchers and students. Two classifications were considered based on available expert annotations. The first classification enables the comparison of sensitive and multi-drug resistance. The second classification compares all of the five aforementioned categories. We demonstrate and discuss comparative prediction statistics for the different algorithms regarding accuracy in the classification and the work required to perform the implementation of the best predictive methods for online resources.

Category: Computational Biology