abc: represent the results of bonferroni comparison,a significant difference between trainee and competent, p<0.05;b significant difference between trainee and expert, p<0.05
Figure 1 Flowchart for the development and test of the algorithms. M&C, Maternal and Child; W&C, Women and Children’s; CNS, central nervous system; AI, artificial intelligence.
Figure 2 Flow chart illustrating the entire process of the network. As shown in the figure, our process contains one input and two outputs. In the first output, two labels were detected on the same side of ventricle by the model, which were lateral ventricle (green box, the label score was 0.597146) and tear-ventricle (lower yellow box, the label score was 0.871927). After label elimination in the logic output network according to the scores, only one label with the higher score remained in output image (tear-ventricle, lower yellow box).
Figure 3 The composite image shows the AI output correctly labeled with corresponding type of specific malformations in each image, as well as normal image. ACC, Absence of corpus callosum; ASP, absence of cavum septi pellucidi; DWNv, Dandy-Walker malformation or variant; HPE, holoprosencephaly; MCM, Megacisterna magna; CPC, choroid plexus cyst.
Figure 4 The performance of the AI system and Ultrasonic doctors in CNS malformations identification a. AI system outperforms the average of the ultrasonic doctors at CNS malformations identification. Each point represented the sensitivity and specificity of a single ultrasonic doctors, the blue points are the average of the doctors, with error bars denoting one standard deviation. The AI system achieves superior performance to a doctor if the sensitivity–specificity point of the lies below the blue curve, which most do. b, The performance of AI model versus that of experts, competent and trainee doctors.
Figure 5 The improvement of overall performance of three degrees of doctors in CNS malformations identification with AI assistance (a. trainee, b. competent, c. expert).