FB-ZWUNet: A deep learning network for corpus callosum segmentation in fetal brain ultrasound images for prenatal diagnostics

Objective: Develop an automated method to replace manual segmentation of the corpus callosum (CC) in fetal brain ultrasound images, thereby reducing reliance on specialist expertise and mitigating human error, ultimately improving CC segmentation accuracy and enhancing fetal brain development assessments. Methods: We propose FB-ZWUNet, an end-to-end CC segmentation network composed of three key modules: (1) Zernike Attention Module (ZAM) for feature enhancement, (2) Wavelet Attention Module (WAM) for optimized feature fusion and precise image reconstruction, and (3) Morphological Constraint Module (MCM) for accurate edge and region capture. The network was trained on the FB-CC dataset, which consists of 1,336 annotated fetal brain mid-sagittal ultrasound images. Results: FB-ZWUNet achieved a Dice coefficient of 0.8743 and IoU of 0.7813, outperforming current state-of-theart methods in accuracy and stability while offering faster inference times. The FB-CC dataset provides comprehensive samples from 18 to 32 weeks of gestation, collected using three different ultrasound devices. Conclusions: FB-ZWUNet effectively addresses segmentation challenges, demonstrating significant clinical potential in assessing fetal neurological development and improving the reliability of early diagnosis and prenatal care.