Evaluating the Impact of Hyperparameters Using U-Net Networks with Transfer-Learning for Cerebellum Segmentation from Ultrasound Scans

The cerebellum is an essential part of the human brain, responsible for motor learning, balance, and coordination. Therefore, cerebellar growth in the fetal brain is identified by its shape on ultrasound images. To overcome the limitations of the current clinical process, several computerized approaches have been proposed in the literature. This research paper proposes cerebellum segmentation from 2D ultrasound images of the fetal brain using U-Net, Attention U-Net, and transfer-learning-based U-Net models, namely VGG16-U-Net, ResNet50-U-Net, and EfficientNetB1-U-Net, with (1) four different evaluation parameters: Dice coefficient, Intersection over Union (IoU), Precision, and loss, (2) three different loss functions, namely Binary Cross-Entropy (BCE) loss, Dice loss, and Tversky loss, (3) three batch sizes – 8, 16, and 32. Lastly, the effectiveness of the trained models for cerebellum segmentation is assessed for real-time data. This research aims to segment the cerebellum in second-trimester fetal brain ultrasound images to predict prenatal neurodevelopmental disorders.