por
Lauren Dubinsky, Senior Reporter | April 28, 2017
a) Chest X-ray showing active TB
b) Same X-ray with heat map overlay
after it was passed through
the GoogLeNet-TA classifier
Artificial intelligence models may soon help screen and evaluate patients with suspected tuberculosis in areas that have limited access to radiologists. Researchers at Thomas Jefferson University Hospital are training the models to identify TB on chest X-rays.
“An artificial intelligence solution that could interpret radiographs for presence of TB in a cost-effective way could expand the reach of early identification and treatment in developing nations," Dr. Paras Lakhani, coauthor of the study, said in a statement.
In a study, Lakhani and his team obtained 1,007 X-rays of patients with and without active TB from the National Institutes of Health, the Belarus Tuberculosis Portal, and TJUH.
Ad Statistics
Times Displayed: 8082
Times Visited: 39 Ampronix, a Top Master Distributor for Sony Medical, provides Sales, Service & Exchanges for Sony Surgical Displays, Printers, & More. Rely on Us for Expert Support Tailored to Your Needs. Email info@ampronix.com or Call 949-273-8000 for Premier Pricing.
The cases were used to train two different deep convolutional neural network models, which resemble brain structure and have multiple hidden layers and patterns to classify images.
The DCNN models, AlexNet and GoogLeNet, learned from both the TB-positive and TB-negative X-rays. They were tested on 150 different cases that weren't used for training and had an accuracy rate of 96 percent.
“The relatively high accuracy of the deep learning models is exciting,” said Lakhani. “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.”
However, the two DCNN models disagreed on 13 out of the 150 test cases. For those cases, the researchers evaluated a workflow in which an expert radiologist interpreted the images, which yielded a greater net accuracy of almost 99 percent.
Going forward, the team plans to make further improvements on the models by deploying more training cases and other deep learning methods.