Deformable attention U-Net network based on fully convolutional discriminator for semantic segmentation of fetal brain ultrasound images

Abstract

Fetal brain ultrasound image segmentation is a crucial task for prenatal diagnosis and medical analysis but remains challenging due to image noise, low contrast, and limited annotated datasets. To overcome these limitations, this study proposes a Deformable Attention U-Net network integrated with a fully convolutional discriminator for semantic segmentation of fetal brain ultrasound images. The proposed approach draws on the generative adversarial network (GAN) concept, where the segmentation model acts as a generator and the discriminator evaluates segmentation quality to optimize network parameters. The segmentation network improves the standard Attention U-Net by introducing 3×3 deformable convolutions to enhance local feature learning and replacing traditional MaxPooling with gating units that selectively retain informative feature maps. This architecture effectively reduces the dependency on large datasets while improving boundary accuracy. Experimental evaluation on a fetal head circumference dataset demonstrated significant performance gains compared to baseline models. The proposed network achieved an Intersection over Union (IOU) of 93.8%, a DICE coefficient of 96.8%, and an accuracy of 97.9%, surpassing U-Net and Attention U-Net by 14.6%, 9.8%, and 4.7%, respectively. These results confirm that integrating deformable attention mechanisms and fully convolutional discriminators enhances edge segmentation and global consistency in fetal ultrasound images. The proposed model provides a reliable framework for fetal biometric measurement and can be further extended to other medical image segmentation applications.