Chest X-Rays with Artificial Intelligence Catches More Lung Cancer

December 13, 2020
Whitney J. Palmer

Using AI with these images significantly improves detection rates, working as a second reader to improve provider performance.

Lung cancer detection and radiologist performance can get a boost from an artificial intelligence (AI) algorithm that pinpoints previously un-detected cancers on chest X-rays.

In a study published in the Dec. 10 Radiology: Cardiothoracic Imaging, investigators from Seoul National University Hospital outlined how a commercially available deep-learning algorithm outperformed four thoracic radiologists on both first and second reads.

Overall, said the team led by Ju Gang Nam, M.D., the algorithm offered both higher sensitivity and higher specificity, and it improved providers’ performance as a seconder reader, leading to significantly improved detection rates.

But, to date, the team said, adoption of computer-aided detection with chest X-ray has been slow because many providers still have lingering questions about whether it can perform well enough in clinical practice. To answer that question, Nam’s team used an enriched dataset of 50 normal chest X-rays, as well as 168 posteroanterior chest X-rays with lung cancers. The average size of the cancers was 2.3 cm +/- 1.2 cm.

For their analysis, they compared the performance of one tool – Insight CXR software from Lunit – to that of the radiologists. The radiologists read the scans twice – once alone and once with the AI tool.

Based on their evaluations, the software yielded better results than the providers. Not only did it produce higher specificity (p=0.01) than the radiologists, but using the tools also enabled higher sensitivity and specificity (p<0.001 and p<0.01, respectively) for the radiologists when they used the software as a second reader. These results were surprising because previous work showed that such tools produced a good number of false positives.

For more coverage based on industry expert insights and research, subscribe to the Diagnostic Imaging e-newsletter here.

It is important to note, the team said, that the providers also failed to reach the performance level the AI tool did when it was used as the sole reader. When reading by themselves, the radiologists overlooked retrocardiac and retrodiaphragmatic nodules. They also mistook true nodules at lower lung fields as nipple or vascular shadow. The difference might occur, they explained, because the AI tool captures pixel values of the image, making it more likely that it will detect abnormalities in soft-tissue density areas that human providers.

But, even with findings that show a higher level of performance, cost will still be an issue around adoption and implementation. In fact, said Charles White, M.D., professor of diagnostic radiology and nuclear medicine at the University of Maryland School of Medicine, it could be the final hurdle due to a lack of reimbursement.

“If this algorithm, indeed, proves far superior to the capability of readers with low false-positive results,” he said, “it would provide a clearer pathway to investing in impetus for further research to determine the precise value of this and other deep learning-based algorithms for nodule detection over chest radiography.”

Related Content:

X-Ray | News