Deep Learning based Fetal Middle Cerebral Artery Segmentation in Large-scale Ultrasound Images

Hemodynamic parameters of the middle cerebral artery (MCA) have great clinical value for diagnosing poor pregnancy outcomes and adverse perinatal results. However, in the process of hemodynamic parameters measurement of MCAs from ultrasound images, sonographers have to manually adjust many different parameters for obtaining a clear color flow image, and then adjust the position and size of the gate to obtain an effective spectral image. To reduce the workload of sonographers, we simplify this procedure by introducing a novel deep learning based system, named as MCANet, with which we can directly obtain the desired position of the gate. In order to implement this system, to our best knowledge, we build the first large-scale MCA dataset consisting of 4005 ultrasound images. Considering that the shape and boundary of MCA determine the most effective position of the gate, we firstly segment the MCA region from the raw ultrasound image and then generate the corresponding position of the gate accordingly. We also propose a novel evaluation metric, Affiliation Index, for measuring the effectiveness of the position of the output gate. Extensive experimental results show that our proposed system outperforms the other state-of-the-art methods in terms of Affiliation Index and all other common evaluation metrics on our proposed dataset.