EP02.49: The anatomical structure analysis in the cross‐vendor fetal cardiac ultrasound videos using deep learning

Abstract

To support examiners in fetal cardiac ultrasound screening, we have developed artificial intelligence (AI) technologies for the automatic detection of cardiac substructures and structural abnormalities in fetal ultrasound videos. However, in clinical practice, these videos are likely acquired from different vendors and hospitals so that the detection performance can drop unexpectedly. In this study, we tried to overcome this domain shift problem toward clinical application. We enrolled 12,026 images from 259 normal fetal cardiac ultrasound videos of 112 cases who were screened in the second trimester. This data set included 3 vendors; Voluson® E8/E10 (GE Healthcare, Chicago, IL, US), Aplio® i700/i800 (Canon Medical Systems, Otawara, Tochigi, Japan), and Arietta® 70 (Fujifilm Healthcare, Tokyo, Japan). We generated fake images like Aplio and Arietta images from Voluson images through the cycle generative adversarial network (CycleGAN) to be used for training the object detection model. The correct positions of 18 different anatomical substructures were annotated with bounding boxes. To evaluate the detection accuracy, the average precision (AP) was calculated and compared to our previously published model trained with only Voluson images. The AP values of several substructures showed a drastic improvement in our proposed/previous model; 0.898/0.451 in the tricuspid valve, 0.846/0.289 in the mitral valve, 0.861/0.677 in the pulmonary artery, 0.854/0.574 in the superior vena cava, 0.793/0.416 in the pulmonary vein, respectively. The mean value of AP was 0.861/0.702. Our proposed model trained with the cross-vendor data set using CycleGAN yielded higher detection performance than the previous model trained with a single-vendor data set. We are potentially enabled to establish a deep learning model using the cross-vendor and limited data set acquired in a clinical scenario.

Ähnliche Arbeiten