Detection and Localization of Carina in X-ray Medical Images with Improved U-Net Model

After tracheal intubation for a patient in the intensive care unit, it is necessary to check for position appropriateness of the intubated endotracheal tube. Timely identification of dislocation and adjustment can prevent patients from morbidity and mortality. Manual checking of the chest X-ray images is time consuming and tedious. An automated way not only speeds the checking but also reduces doctor's work load. In this study, we propose a deep learning model U^(2+)-Net, which yields good performance in semantic segmentation of tracheal and facilitates subsequent localization of the carina. In addition, an algorithm is proposed which locates the coordinate of carina from the segmented trachea. Experimental results show that the overall average error distance of detecting the position of carina is 0.29 cm, accuracy of the detection error within 0.5 cm and 1.0 cm are 85% and 99%, respectively, indicating that the proposed method is promising.