Introduction
Congenital malformations are the leading cause of fetal loss and one of the top ten causes of mortality in children under five1, 2. It also accounted for 25-38 million disability-adjusted life-years worldwide3, which causes heavy burden on individuals, families, health-care systems, and societies4. There are substantial inter-country differences worldwide in the reported prevalence of congenital malformations partly due to the unequal capacities of prenatal screening, leaving many cases undetected, especially in underdeveloped regions. For example, the reported prevalence of congenital cerebral anomalies in Europe increased by 2.4% per annum, but a six-fold difference was found in prevalence across different regions, with an association between prevalence and prenatal detection rate5. Therefore, early identification of congenital anomalies with efficiency is crucial in ensuring medical intervention, minimizing world healthcare disparity, and eventually leading to the optimization of healthcare resources. This goal calls for not only the detection equipment but also doctor expertise for prenatal diagnosis. Yet, training doctors is a timely and costly process, which causes enormous expense to provide prenatal surveillance for average citizens all over the world.
The implementation of artificial intelligence (AI) systems has shown its potential to revolutionize disease diagnosis by performing classification difficult for human experts 6-11. The performance of most reported AI shows a promising trend12-18, furthermore, it has significant advantages in terms of convenient open-source sharing, which have the potential to provide medical guidance to multiple hospitals simultaneously, especially for less developed and remote areas 19,20. In the field of fetal congenital malformation diagnosis, AI development involved the differentiation of images of normal and abnormal fetuses was rare, only limited progress in AI-assisted fetal ultrasound identification of normal fetus structure were reported14-18 , these studies laid a foundation for the development of AI system to identify abnormal structure in ultrasound images by training on fetuses with congenital malformation.
We have initially constructed an AI system involving abnormal fetal CNS ultrasound images to classify fetal CNS ultrasound images as either normal or abnormal and our system achieved a high performance21. Nonetheless, this system only classified images to provide binary outcomes, it is far from making diagnosis for specific CNS malformation. Here, we sought to further advance our system from binary classification to multi-classification, which is capable of detecting multiple types of CNS malformations. We also assessed the efficacy of this algorithm in improving clinical doctors’ diagnostic performance. This is so far the first attempt to construct a deep learning AI system to aid both the experienced and unexperienced physicians in the prenatal ultrasound diagnosis on congenital anomalies.