Fetal Head Segmentation and Head Circumference Measurement Using Residual U-Net

Fetal head segmentation plays an important role in monitoring fetal growth during pregnancy. Automating the segmentation of the fetal head contributes to reducing the workload of healthcare professionals and ensures more consistent and reliable assessments of fetal development. This study introduces a deep-learning-based approach Residual U-Net for fetal head segmentation from ultrasound images and a complete framework for fetal growth monitoring. Residual U-Net utilizes residual connection in its architecture lowering the possibility of vanishing gradients. In the preprocessing stages, contour filling of ground truth, image resizing, image normalization, and data augmentation on an open-source dataset HC18 Grand Challenge was applied. We trained Residual U-Net, U-Net, and Attention U-Net on the same dataset. Evaluation was done on these models based on Dice Coefficient, Jaccard Index, Orientation Angle Difference (OAD), Euclidean Distance (ED), and MSE. The results of the comparative analysis of the models indicated that the Residual U-Net performed better, with an average Dice Coefficient of 97.17% and a Jaccard Index of 94.51%. The performance metrics OAD and ED have average values of 8.29 degrees and 9.58 mm respectively, indicating better Residual U-Net performance than the other models. Also, the calculation of fetal head orientation, and head circumference measurements followed by the calculation of major axis length, minor axis length, and orientation angle was done using ellipse curve fitting.