Automated Segmentation for Numbering Thoracic Vertebrae in Chest Radiographs in Newborn Infants: A Deep Learning Approach

초록

Background and Aims: Identification and labeling of thoracic vertebrae in chest radiographs are widely used to ensure the proper placement of tubes and catheters, as well as to verify the inhalation status in critically ill newborn infants in NICUs. We introduce a fully automated deep-learning system designed to accurately assign numbers to thoracic vertebrae in chest radiographic images. Methods: Portable supine chest X-ray images were prospectively collected from the NICUs of 10 university hospitals in Korea. Ground truth of manual segmentations for the 1st, 7th, and 12th thoracic vertebrae was generated by five radiological technologists and subsequently validated by two neonatologists. The dataset was divided into Train, Validation, and Test sets, with 11,860, 1,400, and 1,400 images, respectively, evenly distributed by GA and BW. The deep-learning model for automatic segmentation was developed using a U-Net. Results: The automatic segmentation algorithm demonstrated excellent agreement with human-annotated segmentation for the T1, T7, and T12 vertebrae (Dice similarity coefficient: 0.8327, 95% CI: 0.8237–0.8418; 0.8322, 95% CI: 0.8213–0.8432; 0.7998, 95% CI: 0.7864–0.8133). For accuracy, at a clinically reasonable Dice similarity coefficient cutoff value of 0.60, the fractions of correct predictions over total predictions for the T1, T7, and T12 vertebrae were very high (95.7%, 94.1%, 90.6%, respectively). Conclusions: Our deep-learning–based automated algorithm precisely detected and categorized the number of thoracic vertebrae on chest radiographs in newborn infants. The artificial intelligence analysis for chest radiographic images will be potential and promising for clinical practice, even as an assistant to medical staff in the NICU.