You are here

Computer-aided TB Screening on Chest X-rays

banner image collage showing 4 pictures: portable x-ray device, mobile x-ray truck, chest-xray, and map of Kenya
Project information

We are collaborating with AMPATH (Academic Model Providing Access to Healthcare), an organization supported by USAID that runs the largest AIDS treatment program in the world. This project uses LHNCBC’s imaging research and system development to fulfill NIH global health policy objectives. Our objective is to leverage in-house expertise in image processing to screen HIV-positive patients in rural Kenya for evidence of pulmonary tuberculosis (TB) in chest x-rays. Since chest radiography is important to the detection of TB and other pulmonary infections prevalent in HIV-positive patients, we have provided AMPATH with lightweight digital x-ray units readily transportable in rural areas. Their staff will take chest x-rays (CXR) of the population and screen them for the presence of disease. These x-ray units are already on site in the Moi University Hospital in Eldoret, Kenya and are being readied for deployment. The team completed design of vehicles to transport the x-ray units – one vehicle is being outfitted as a mobile x-ray truck.

Since the lack of sufficient radiological services in the area suggests the utility of automation to perform the screening, our in-house research effort focuses on developing software to automatically screen for disease in the CXR images. Our researchers are developing algorithms to automatically segment the lungs, detect and remove ribs, heart, aorta and other structures and then detect texture features characteristic of pulmonary disease,. At present, the algorithms distinguish between 2 cases: abnormal vs. normal. After receiving IRB exemption, we obtained chest x-ray images to use as test and training sets: 400 from Montgomery County’s TB Control Program, 850 from a source in India, and 8,200 from Indiana University. We also acquired an open-source Japanese set containing about 250 x-ray images. We are in the process of receiving a few hundred images from China.

Using these x-rays for training and testing, we have developed algorithms for detecting lungs and ribs focusing on region-based features such as log Gabor wavelets that exploit the orientation of anatomical structures. For lung segmentation, we have developed algorithms using region-based level sets and a novel graph-cut segmentation method, yielding an accuracy of about 95%. A robust identification of lung shape plays a role in detecting TB in CXR since many abnormalities (e.g., pleural effusion) exhibit deformation in lung shape. After extracting the lung fields, the algorithm measures various geometric features that discriminate between normal and effused cases. Ongoing work is in identifying the most successful geometric features.

In parallel, we are developing a binary support vector machine (SVM) classifier that uses several features extracted from the x-rays as input, such as histograms of intensity, gradient magnitude and orientation, shape and curvature. Based on these input features, the SVM returns a confidence value, allowing an operator to inspect cases in which the classifier is uncertain. This initial classifier, showing an accuracy of about 80%, serves as our starting point for ongoing optimization.

Publications/Tools: 
Candemir S, Rajaraman S, Thoma GR, Antani SK. Deep Learning for Grading Cardiomegaly Severity in Chest X-rays: An Investigation. Proc. IEEE Life Sciences Conference (LSC 2018), Montreal, Quebec, Canada, 28 – 30 October 2018. pp. 109-113.
Rajaraman S, Candemir S, Kim I, Thoma GR, Antani SK. Visualization and Interpretation of Convolutional Neural Network Predictions in Detecting Pneumonia in Pediatric Chest Radiographs. Appl. Sci. 2018, 8, 1715
Jaeger S, Juarez-Espinosa O, Candemir S, Poostchi Mohammadabadi M, Yang F, Kim L, Ding M, Folio L, Antani SK, Gabrielian A, Hurt D, Rosenthal A, Thoma GR. Detecting drug-resistant tuberculosis in chest radiographs International Journal of Computer Assisted Radiology and Surgery https://doi.org/10.1007/s11548-018-1857-9
Vajda S, Karargyris A, Jaeger S, Santosh KC, Candemir S, Xue Z, Antani SK, Thoma GR. Feature Selection for Automatic Tuberculosis Screening in Frontal Chest Radiographs. J Med Syst. 2018 Jun 29;42(8):146. doi: 10.1007/s10916-018-0991-9.
Rajaraman S, Candemir S, Xue Z, Alderson P, Kohli M, Abuya J, Thoma GR, Antani SK. A novel stacked generalization of models for improved TB detection in chest radiographs. Proc. IEEE Engineering in Medicine and Biology Conference (EMBC 2018), Honolulu, Hawaii, 2018. pp. 718-721.
Xue Z, Long LR, Jaeger S, Folio L, Thoma GR. Extraction of Aortic Knuckle Contour in Chest Radiographs Using Deep Learning. EMBC 2018.
Zohora FT, Antani SK, Santosh KC. Circle-like foreign element detection in chest x-rays using normalized cross-correlation and unsupervised clustering. Proc. SPIE 10574, Medical Imaging 2018: Image Processing, 105741V (2 March 2018); doi: 10.1117/12.2293739; doi.org/10.1117/12.2293739.
Rajaraman S, Antani SK, Candemir S, Xue Z, Abuya J, Kohli M, Alderson P, Thoma GR. Comparing deep learning models for population screening using chest radiography. Proc. SPIE 10575, Medical Imaging 2018: Computer-Aided Diagnosis, 105751E (27 February 2018).
Xue Z, Antani SK, Long LR, Thoma GR. Using deep learning for detecting gender in adult chest radiographs. SPIE Medical Imaging 2018
Xue Z, Jaeger S, Antani SK, Long LR, Karargyris A, Siegelman J, Folio LR, Thoma GR. Localizing tuberculosis in chest radiographs with deep learning. SPIE Medical Imaging 2018

Pages