Chest radiography is the most commonly performed radiologic examination, but interpretation of these studies isn't easy: 22% of reading errors in diagnostic radiology were made on chest x-rays, said senior study author Dr. Chang Min Park, PhD, of Seoul National University College of Medicine. In addition, the growth rate of chest radiography has been much faster than the increase in the number of qualified radiologists.
To help, the researchers developed a deep learning-based detection algorithm for chest radiographs that would cover several major thoracic diseases, including pulmonary malignancies, active pulmonary tuberculosis (TB), pneumonia, and pneumothorax. They found that their algorithm showed high performance in external validation tests and also did better than physicians -- including thoracic radiologists -- in classifying images and localizing lesions. The algorithm also enhanced physicians' performance when used as a second reader.
The results show that the algorithm has promise as a standalone interpreter of chest radiographs in selected clinical situations, such as triaging a large volume of screening radiographs for TB screening or serving as a first reader prior to a confirmative reading by radiologists in areas of poor medical resources, according to the researchers. What's more, it could help clinical workflow by prioritizing chest radiographs with suspicious findings that would require prompt diagnosis and management, Park noted.
"In addition, improved performances of physicians with assistance of our algorithm can indicate the potential of our algorithm as a second reader to improve the quality of [chest radiography] reading, particularly of physicians with less experience," Park told AuntMinnie.com. "Our algorithm can be used to improve the quality of interpretation of [chest radiographs], especially in situations where expert radiologists are not available."
Attend this presentation by first author Eui Jin Wang on Wednesday to learn more.
