Automatic Measurement of Fetal Cavum Septum Pellucidum From Ultrasound Images Using Deep Attention Network
Yuzhou Wu, Kuifang Shen, Zhigang Chen, Jia Wu
Abstract
The measurement of cavum septum pellucidum is an important step in prenatal testing. However, this process is usually done manually, which is such a difficult and time-consuming task due to the attenuation and shadows of ultrasound images even for experienced sonographers. In this study, we propose a novel deep attention network to address this problem by segmenting and measuring the width of cavum septum pellucidum. The proposed network is based on U-net with three changes: a new channel attention module, increasing attention on relevant regions; VGGI I, adding the depth of encoder path to increase the receptive field; And post-processing to measure and diagnose the anomalies of cavum septum pellucidum. Experiments on a fetal ultrasound dataset demonstrated our proposed network achieved the highest precision of 79.5% and the largest Dice score of 77.5%. To demonstrate the generalization capacity, we also have been validated our model on the BraTs 2017 dataset, obtaining an excellent performance with the Dice score of 91.5%.