A tuberculosis blood or skin test is limited in detecting the progress of the infection, and there is a lack of radiologists with the expertise to identify and diagnose TB on chest images. However, all of that is set to change since a recent study published in Radiology found a way to apply deep learning with x-ray interpretation to sharpen the accuracy of TB detection.
According to the World Health Organization, TB is considered to be one of the top 10 causes of death. In 2015, 10.4 million people contracted the disease and 1.8 million were unable to survive it. TB is prevalent in developing countries such as India, China, Indonesia, Nigeria, and Pakistan, but is growing exceedingly less common in the United States — with only 9,421 reported cases in 2014, according to the Centers for Disease Control and Prevention. TB can be identified via x-ray images, however the number of radiologists with the training and skills to accurately diagnose the disease are few and far between.
That’s where the team of researchers from Thomas Jefferson University Hospital (TJUH) in Philadelphia come in. The group led by Paras Lakhani, MD, examined 1,007 x-rays provided by the National Institute of Health, the Belarus Tuberculosis Portal, and TJUH of patients both with and without active TB. The researchers divided the cases into three categories: training (68 percent), validation (17.1 percent), and test (14.9 percent). The cases were used to train two different DCNN models, AlexNet and GoogLeNet, to differentiate between TB-positive and TB-negative images. They found that the best artificial intelligence model was a blend of the two models, which performed at a net accuracy of 96 percent.
“The relatively high accuracy of the deep learning models is exciting,” Dr. Lakhani told the Radiological Society of North America. “The applicability for TB is important because it’s a condition for which we have treatment options. It’s a problem that can be solved.”