Automatic Extraction Of Fetal Corpus Callosum And Cerebellar Vermis Ultrasound Images

In our paper,we present an integrated framework for automatic localization and segmentation of the ‘Complex of Corpus Callosum and Cavum Septum Pellucidum(CC&CSP)’ and ‘Cerebellar Vermis(CV)’ in ultrasound images of the fetal brain based on a priori anatomical knowledge and computer vision methods.The framework aims to obtain area and location information of CC&CSP and CV in median plane of fetal brain to assist physicians in prenatal screening examination.There has been no study specially addressing this issue and this study has important clinical implications for early diagnosis of human embryos.The framework proposed in this paper is implemented as follows.First,several CC&CSP and CV local images are manually cut from the median plane of fetal brain,and average templates of both are generated by Variational Autoencoder(VAE).Second,the CC&CSP average template is used to locate the CC&CSP in ultrasound images using the "initial localization-accurate localization-result detection" strategy with step-by-step accuracy.Third,using the morphological characteristics of CC&CSP,we perform "initial contour fitting-contour iteration" for the segmentation of CC&CSP.Fourth,using the relative position of CC&CSP and CV,the CV can be localized in the same way as the CC&CSP localization.Finally,the initial contour fitting is performed to complete the segmentation of CV.Through this serial and integrated structure,the localization and segmentation can be accomplished by taking full advantage of the morphological characteristics and interrelationships of the target organs.The framework allows for efficient and accurate localization and segmentation of the CC&CSP and CV for multiple types of experiments.Experimental results proved the accuracy and reliability of this method.The method of this paper is currently being advanced to clinical trials in Shengjing Hospital of China Medical University.