By Will Boggs MD
NEW YORK (Reuters Health) - 29/3/2019
A new deep-learning-based automated-detection (DLAD) algorithm accurately classifies major thoracic diseases on chest x-rays, researchers from South Korea report.
"Chest radiograph is the most frequently performed radiologic examination in daily practice, and its interpretation is important in diagnosis and management for many patients; however, it is not an easy task," said Dr. Chang Min Park from Seoul National University College of Medicine.
"The key message of our paper is that it is possible to identify major thoracic diseases on chest radiographs at the level of expert subspecialty radiologists with the deep-learning algorithm, and that if properly utilized, it would improve the quality and efficiency of our clinical practice, and hopefully clinical outcomes," he told Reuters Health by email.
Dr. Park and colleagues in the DLAD Development and Evaluation Group earlier investigated DLAD algorithms for classifying chest x-rays with malignant nodules and active pulmonary tuberculosis,. In the current report, they describe their development of a DLAD algorithm for major thoracic diseases on chest radiographs and the validation of its performance in comparison with physicians. The major thoracic diseases targeted by the DLAD algorithm included pulmonary malignant neoplasms (including primary lung cancers and metastases), active pulmonary tuberculosis, pneumonia and pneumothorax. In the in-house validation data set, the DLAD algorithm showed 96.5% accuracy (based on AUROC) for differentiating normal from abnormal chest x-rays.
For external validation, the algorithm showed 97.9% accuracy, the team reports in JAMA Network Open, online March 22.
The DLAD algorithm also accurately localized the included major thoracic diseases imaged on chest radiograph (91.6% in-house validation and 97.2% external validation).
By comparison, the pooled AUROCs for differentiating abnormal from normal chest x-rays were 81.4% for non-radiology physicians, 89.6% for board-certified radiologists, and 93.2% for thoracic radiologists, significantly lower than the 98.3% for the DLAD algorithm. There was a similar pattern for lesion-wise localization.
The DLAD algorithm had variable accuracy for differentiating the individual thoracic diseases from normal findings: 84.0% for pulmonary malignant neoplasm, 20.9% for active pulmonary tuberculosis,, 73.1% for pneumonia and 95.0% for pneumothorax. Dr. Park mentioned several possible applications of the algorithm.
"First, this algorithm can prioritize work-list of radiographs according to the presence of clinically relevant diseases, which can improve quality of radiology practice and reduce turnaround time of patients with clinically relevant diseases," he said.
"Second, it can improve the accuracy of interpretation in the daily practice by assisting interpreting physicians, especially less-experienced physicians. Finally, in a selected situation where an interpreting physician is not available, it can be utilized as a standalone diagnostic tool to classify chest radiographs with major thoracic diseases," Dr. Park said.
"Deep-learning technology has a great potential and has already contributed in the medical field," he added. "However, it is not a panacea. The current deep learning technology has its own weaknesses. We need to know them in order to properly utilize this wonderful technology. In a field like medicine, meticulous validation is really important in every step."
JAMA Netw Open 2019.