VP18.04: Automatic evaluation of the cardiothoracic area ratio and cardiac axis in fetal ultrasound using deep learning

• DOI: 10.1002/uog.24300
• Publication Date: 2021-10-18T19:14:26Z
• Authors: K. Shozu; M. Komatsu; A. Sakai; A. Dozen; R. Komatsu; T. Arakaki; S. Yasutomi; R. Matsuoka; A. Nakashima; A. Sekizawa; R. Hamamoto
• Keywords: Vertebra; Fetal heart

The abnormal findings of the fetal cardiothoracic area ratio (CTAR) and cardiac axis (CA) are manifested as a result various factors such heart dysfunction compensatory state. Therefore, these indexes convenient highly useful for primary screening. We have been developing artificial intelligence (AI) powered diagnostic support technologies ultrasound. In this study, we constructed deep learning-based model that automatically measures above in cross-section with four-chamber view (4CV) evaluated its performance. enrolled 1142 4CV ultrasound images 279 normal cases (the range gestational age: 18–28 weeks). These were annotated ground truth labels heart, lung, vertebra, ventricular septum. performed segmentation using DoubleU-net, processed each prediction label, combined necessary information to construct our model. calculated label pixels/ thoracic pixels CTAR. also angle CA, which was formed between approximate straight line septal passed through vertebra dividing equally thorax into two parts. made performance evaluation 50 25 test dataset. results septum values intersection over union (IoU), 0.887, 0.976, 0.706, 0.662, respectively. mean absolute error (MAE) CTAR 1.94%, CA 10.1°. It suggested proposed could calculate by inputting combining structure. would like further determine appropriate image depiction verify operation cases.