Detection And Diagnosis Of Craniocerebral Anatomical Structure In Ultrasound Based On Deep Learning Method

Central nervous system abnormalities are some of the most common congenital malformations that may associated with syndromic and chromosomal malformations,could lead to neurodevelopmental delay and mental retardation,and bring great mental suffering and financial burden to the patients’ family.Transabdominal ultrasound is one of the routine methods for imaging the fetal brain anatomies to screen the central nervous system abnormalities.However,this is a challenging task to acquire and evaluate the fetal brain anatomies,which usually require long-time knowledge training and experience accumulation.In most regions and countries,the lack of professional ultrasound analysts makes the severe central nervous system abnormally prone to miss detection in early prenatal screening,which leading to a series of adverse consequences.Aiming at diagnosis for central nervous system abnormalities,this work introduces the latest deep learning technic to design efficient and automatic method for detection and diagnosis of the fetal craniocerebral structure,which is expected to reduce the shortage of medical resources in economically underdeveloped areas.The main work and contributions are as follows:(1)We propose an anomaly detection method for ultrasonic craniocerebral image via row and column information classification.The framework train and validate on a large clinical dataset of about 20,000 standard craniocerebral ultrasound plane images containing 7common fetal brain abnormalities.The experiment results show that the framework robustly segment the craniocerebral regions,accurately classify fetal brain images of standard neurosonographic planes as normal or abnormal and locate the lesions to provide visualized evidence for the diagnosis,which may provide assistance to doctors for brain abnormalities detection in antenatal neurosonographic assessment.(2)We propose a real-time detection network with temporal information fusion to detect the craniocerebral structures in ultrasonic video.We used the data obtained from the realworld clinical routines to train and validate the performance of network.The experiment results show that temporal module play an important role to enhance the detection result.Thought optimization,the proposed detection network not only meets the real-time clinical needs,but also achieves the comparable performance with the state-of-arts methods.