Radiologists Spot Chest X-ray Abnormalities Better with Deep-Learning Detection

Using a deep-learning detection system at the time of image interpretation – not as a second reader – improves accuracy.

Simultaneous use of a deep-learning based detection (DLD) system improves a radiologist’s accuracy in identifying major abnormalities on chest X-rays, a new study has found.

While the efficacy of artificial intelligence (AI) algorithms as a second reader has been well established in previous studies, providers included in those investigations have read scans sequentially. Interpreting images with computer-aided detection (CAD) and, then, without the tool introduces both reading order and recall bias, said a team from South Korea.

Instead, they said, radiologists were faster and more accurate with pinpointing abnormalities when they used CAD. The team published the results of their retrospective, randomized trial on March 23 in Radiology.

“In the rigorous setting of our study, which minimized bias, the diagnostic performance of the radiologists was substantially improved with the assistance of the DLD system,” said the team led by Jinkyeong Sung from the radiology department at the University of Ulsan College of Medicine, Asan Medical Center in Seoul, South Korea.

For their study, the team collected 114 abnormal and 114 normal chest X-rays captured between January 2016 and December 2017. A group of six providers, including thoracic radiologists, evaluated the scans both with and without the DLD system, using a cross-over design and a washout period.

According to the team’s analysis, the radiologists performed better when using the DLD. Not only did their per-lesion sensitivity rise from 83 percent to 89.1 percent, but their per-image sensitivity also rose from 80 percent to 89 percent. Specificity also improved from 89.3 percent to 96.6 percent.

In addition, localization increased from 0.90 to 0.95, and the area under the receive operating curve (AUC) jumped from 0.93 to 0.98. Using the DLD also decreased reading time from 10-to-65 seconds down to 6-to-27 seconds.

The team also noted that, when used alone, the DLD outperformed the pooled observers with localization of 0.96 compared to 0.90 and AUC of 0.98 compared to 0.93.

Even with these results, the team said, additional research is needed.

“Further studies in other institutions and countries are needed to ensure generalizability,” they said. “In addition, future studies in the real-world setting will help determine the usefulness of the DLD system.”

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