Researchers developed an automated algorithm that was able to detect active pulmonary tuberculosis on chest radiographs better than all groups of physicians, including thoracic radiologists.
They said it could improve diagnosis and screening of TB in certain situations and reduce the high global burden of the disease.
“Detection of active pulmonary TB on chest radiographs (CRs) is critical for the diagnosis and screening of TB,” Chang Min Park, MD, PhD, associate professor of diagnostic radiology at Seoul National University College of Medicine, and colleagues wrote. “An automated system may help streamline the TB screening process and improve diagnostic performance.”
The researchers used 54,221 normal CRs and 6,768 CRs with active pulmonary TB — all labeled and annotated by board-certified radiologists — to create a deep-learning-based automatic detection (DLAD) algorithm that they tested using six independent datasets, including CRs from China, South Korea and the United States. They compared its performance agains 15 physicians with varying degrees of experience, including thoracic radiologists, board-certified radiologists and nonradiology physicians.
According to evaluation metrics, the algorithm performed at a consistently high level, demonstrating classification performances between .977 and 1 and localization performances ranging from .973 to 1, with sensitivities and specificities of classification between 94.3% to 100% and 91.1% to 100%, respectively, using the high-sensitivity cutoff and 84.1% to 99% and 99.1% to 100% using the high-specificity cutoff, the researchers reported. It outperformed all physicians in both measures — a first for an automated algorithm, they said.
The researchers also validated the potential of the algorithm as a second reader and found that nonradiology physicians assisted by DLAD showed improvements in both sensitivity and specificity, and board-certified radiologists showed improvement in sensitivity.
“In conclusion, our DLAD demonstrated excellent and consistent performance in the detection of active pulmonary TB on CR, outperforming most physicians including thoracic radiologists,” the authors concluded.
“For clinical application of our DLAD, two scenarios can be considered. First, our DLAD may have the potential as a second reader in clinical practice, which would improve the performance of physicians for active TB, especially in primary healthcare or community-based settings where interpretation of CRs should be done by primary care providers rather than expert radiologists. Second, high performance of our DLAD in classifying TB-CRs, outperforming even thoracic radiologists, may suggest the potential of the stand-alone utilization of DLAD in screening patients with active TB or in triaging CRs requiring experts’ reading.” – by Caitlyn Stulpin
Disclosures: Park reports grants from Seoul National University Hospital. Please see the study for all other authors’ relevant financial disclosures.