CNSeg-GAN: A Lightweight Generative Adversarial Network For Segmentation of CRL and NT From First-Trimester Fetal Ultrasound

This paper presents a novel, low-compute and efficient generative adversarial network (GAN) design for automatic segmentation called CNSeg-GAN, which combines 1-D kernel factorized networks, spatial and channel attention, and multi-scale aggregation mechanisms in a conditional GAN (cGAN) fashion. The proposed CNSeg-GAN architecture is trained and tested on a first-trimester ultrasound (US) scan video dataset for automatic detection and segmentation of anatomical structures in the midsagittal plane to enable Crown Rump Length (CRL) and Nuchal Translucency (NT) measurement. Experimental results shows that the proposed CNSeg-GAN is x15 faster than U-Net and yields mIoU of 78.20% on the CRL and 89.03% on the NT dataset, respectively with only 2.19 millions in parameters. The accuracy of this lightweight design makes it well-suited for real-time deployment in future work.