Automating cardiothoracic ratio measurements in chest X-rays

Abstract

The analysis of the positions, shapes, and sizes of thoracic organs is an internationally established practice for radiologists. The considerable amount of time spent on manual measurements of roentgenographic features reveals the need for a computerized approach for the automation of these measurements. In this work, we introduce a new way for the annotation of the chest x-ray data and evaluation of the most commonly-calculated morphometric parameter - cardiothoracic ratio. The measurement of interest was defined as ratio of line segments outlining the heart and the distance between two most lateral landmarks on lung fields. Using a manually annotated dataset, we developed a hourglass-based deep learning-based model to detect these landmarks and perform the measurement. We found that the predictions of the proposed solution differ from the annotation of an expert radiologist in 9.8mm error measured in terms of the mean Euclidean distance.