The purpose of this scientific work is to conduct a comparative analysis of the diagnostic effectiveness, time, and economic indicators of deep learning algorithms and expert assessment in the diagnosis of brain development abnormalities in newborns according to neuroimaging data from the UNICEF BIGH dataset. In a retrospective study, 300 ultrasound and MRI scans were used. Each case was independently analyzed by the U-Net segmentation model, the EfficientNet classification model, a group of undergraduate students, and an expert doctor. The diagnostic accuracy (sensitivity, specificity, F1-score, Dice coefficient), average analysis time, and estimated cost for 100 cases were estimated. The EfficientNet model demonstrated 97% sensitivity in the binary classification task "norm/pathology", showing a result comparable to the expert. The U-Net model achieved the accuracy of segmentation of the ventricles of the brain with a Dice coefficient of 0.92. The analysis time of the combined AI pipeline was 1.8 minutes per case, compared to 13 minutes for an expert and 24 minutes for a group of students. The estimated cost of analyzing 100 cases by an expert was 1085 USD, while the cost of using an AI pipeline was 15 USD. Therefore, by combining the U-Net and EfficientNet algorithms, it is possible to reach diagnostic accuracy that is on par with that of an expert while saving a significant amount of time and money. The most effective implementation model seems to be a two-stage system in which AI performs primary screening and quantification, and an expert verifies complex cases.
