Deep learning-based cardiothoracic ratio measurement on chest radiograph: accuracy improvement without...

A reproducible and accurate automated approach to measuring cardiothoracic ratio on chest radiographs is warranted. This study aimed to develop a deep learning-based model for estimating the cardiothoracic ratio on chest radiographs without requiring self-annotation and to compare its results with those of manual measurements.

The U-net architecture was designed to segment the right and left lungs and the cardiac shadow, from chest radiographs. The cardiothoracic ratio was then calculated using these labels by a mathematical algorithm. The initial model of deep learning-based cardiothoracic ratio measurement was developed using open-source 247 chest radiographs that had already been annotated. The advanced model was developed using a training dataset of 729 original chest radiographs, the labels of which were generated by the initial model and then screened. The cardiothoracic ratio of the two models was estimated in an independent test set of 120 original cases, and the results were compared to those obtained through manual measurement by four radiologists and the image-reading reports.

The means and standard deviations of the cardiothoracic ratio were 52.4% and 9.8% for the initial model, 51.0% and 9.3% for the advanced model, and 49.8% and 9.4% for the total of four manual measurements, respectively. The intraclass correlation coefficients (ICCs) of the cardiothoracic ratio ranged from 0.91 to 0.93 between the advanced model and the manual measurements, whereas those for the initial model and the manual measurements ranged from 0.77 to 0.82.

Deep learning-based cardiothoracic ratio estimation on chest radiographs correlated favorably with the results obtained through manual measurements by radiologists. When the model was trained on additional local images generated by the initial model, the correlation with manual measurement improved even more than the initial model alone.