Automated Carina Detection in Chest X-ray Images Using Non-Overlapping and Cross-Squeeze Convolutional Neural Networks

A chest X-ray test is necessary to verify the correct positioning of several medical devices and to check for any issues that may have resulted from improper placement. The trachea, carina, and bronchi localization in chest radiography is a crucial test for patients in the Intensive Care Unit (ICU). Most of the images become affected by the machine or the shooting camera's position, so a physician is required to manually adjust the images to locate Carina. Therefore, this research aim is to automate Carina detection in chest X-ray images without manual adjustment. We proposed a Non-Overlapping Max Pooling (NOMP) connection and Cross-Squeeze Convolutional Attention (CSCA) to make up the design architecture. In addition, a convexity defect technique is combined with the post-processing algorithm to determine Carina's final placement. The experimental results show that the overall performance for average error distance is 0.3097 cm, the accuracy rate is 81.5% when the error falls within 0.5 cm, and the accuracy rate is 97% when the error falls within 1 cm.