Integrated generation adversarial and semi-supervised network for Corpus Callosum and Cavum Septum Pellucidum Complex segmentation in fetal brain ultrasound via progressive training

Abstract

In prenatal diagnostics, measuring the Corpus Callosum (CC) is crucial for assessing fetal brain development, with the Cavum Septum Pellucidum (CSP) also playing a key role due to its close association with the CC. Our study addresses the challenge of distinguishing these interconnected structures by segmenting the Corpus Callosum and Cavum Septum Pellucidum Complex (CCSP) as a unified entity in mid-sagittal fetal brain ultrasound images. This approach ensures accurate biological measurements that reflect their combined significance in brain development. To improve segmentation accuracy and reduce errors inherent in manual methods, we propose the Fetal Brian Ultrasound Semi-supervised Generative Adversarial Segmentation Network (FB-SGASNet) tailored for few-shot datasets. FB-SGASNet enhances clinical applicability through: (i) an integrated framework combining the Target Segmentation Module (TSM) and Data Expansion Module (DEM) with a semi-supervised learning strategy; (ii) a progressive training strategy promoting parameter sharing between TSM and DEM; (iii) the introduction of Feature Fusion Attention Module (FFAM) and Dual-Stream Feature Attention Module (DSFAM) to improve key anatomical feature recognition; and (iv) the use of a specialized Fetal Brain CCSP dataset (FB-CCSP) with 200 annotated images for network training and validation. FB-SGASNet provides a practical solution for CCSP segmentation in few-shot datasets, enhancing the efficiency and accuracy of fetal brain ultrasound analysis, reducing reliance on specialized expertise, and enabling more timely prenatal evaluations.