Chest X-Ray Analysis Can Predict Pulmonary-to-Systemic-Flow Ratio

By Will Boggs MD

January 28, 2020

NEW YORK (Reuters Health) - A deep-learning-based analysis of chest radiographs accurately predicts high pulmonary-to-systemic-flow ratios in patients with congenital heart disease (CHD), researchers in Japan report.

"This deep learning-based system will confer an opportunity for quantitative and objective analysis of pulmonary-blood-flow status in chest radiographs in the daily practice," said Dr. Shuhei Toba of Mie University Graduate School of Medicine, in Tsu, and Boston Children's Hospital, in Boston.

"In addition to the superb performance of the present system, deep learning may give us a novel medical insight into the chest-radiograph findings predictive of hemodynamic status, which were otherwise unrecognized as such," he told Reuters Health by email.

The pulmonary-to-systemic-flow ratio is an important hemodynamic parameter for clinical decision-making in patients with CHD, but its measurement currently requires an invasive procedure. Deep learning has been used to analyze radiographic images, but it has not yet been applied to analyzing chest radiographs in these patients.

To develop and validate their deep-learning-based method, Dr. Toba and colleagues used Inception-v3, a deep convolutional neural network model developed by Google.

The mean Fick-derived pulmonary-to-systemic-flow ratio from 1,031 catheterizations performed for 657 patients was 1.43 and was 2.0 or more in 18% of catheterizations in the training group and 19% of catheterizations in the evaluation group.

The intraclass correlation coefficient for the Fick-derived and deep-learning-derived pulmonary-to-systemic-flow ratio was 0.68, the researchers report in JAMA Cardiology.

The diagnostic concordance rate between deep learning-derived and Fick-derived classification was 64% (64/100), which was significantly higher than that of certified pediatric cardiologists (49%) or pediatric cardiology fellows (40%).

The deep-learning-derived model detected a high ratio of 2.0 or more with 86% accuracy, compared to 80% for pediatric cardiologists and 78% for pediatric cardiology fellows.

According to the model, the area in the lung fields (coarse nodules, in particular) and the area around the heart were valuable for predicting an increased ratio, whereas no particular focus pattern predicted a decreased ratio.

"We believe that the deep-learning-based model (or artificial intelligence) will be most useful for screening or outpatient clinic, where specialists are not readily accessible," Dr. Toba said. "In addition, if the performance of our model is further optimized by teaching (it) more cases, it will surely be used in the daily practice to evaluate pulmonary-blood-flow status, just like a very common parameter, the cardiac thoracic ratio, for cardiomegaly in chest radiographs."

The study had no commercial funding, and the researchers report no conflicts of interest.

SOURCE: JAMA Cardiology, online January 22, 2020.