A novel technology based on deep learning for the identification and classification of fetal structures in the first trimester

As technologic advances of ultrasound with high resolution, the importance and accuracy of identifying fetal structural abnormalities in first trimester has increased. We aimed to develop an artificial intelligence system called ViewAssist™ that automatically identifies and classifies the fetal structures in the first trimester and to evaluate its performance. Ultrasound images of over 8,000 first trimester fetuses were prospectively collected from Seoul St. Mary's Hospital, Eunpyeong St. Mary's Hospital, Uijeongbu St. Mary's Hospital, and Buchoen St. Mary's Hospital. The collected images were classified by anatomical areas including head, face, neck, chest, heart, abdomen, extremities, and spine, according to the International Society of Ultrasound in Obstetrics and Gynecology guideline. For effective classification of the anatomical areas, we used the Vision Transformer architecture and pre-processed the data by augmenting images while maintaining the aspect ratio. We randomly selected images from the original data, with 90% for the training set and 10% for the validation set. During the deep learning training stage, we used 100 epochs, with a batch size of 64 and a learning rate of 0.0005. We adjusted the hyperparameters at each epoch to minimise the loss function. The classification accuracy of the anatomical areas using ViewAssist™ was 96.75% overall, with the following category-specific accuracy: 100% for head, 98.9% for face, 93.8% for neck, 85.7% for chest, 100% for heart, 94.5% for abdomen, 97.5% for extremities, and 96.2% for spine. Through deep learning algorithms, ViewAssist™ was able to accurately classify the structures of first trimester fetuses. It is expected to be clinically useful in assisting with identifying fetal structures and diagnosing structural abnormalities in the first trimester.