Generating Synthetic B-Mode Fetal Ultrasound Images Using CycleGAN-Based Deep Learning

B-mode ultrasound (USG) is a key imaging modality for fetal assessment, providing a noninvasive approach to monitor anatomical development and detect congenital anomalies at an early stage. However, portable ultrasound devices commonly used in low-resource healthcare settings often yield low-resolution images with significant speckle noise, reducing diagnostic accuracy. Furthermore, the scarcity of labeled medical data, caused by privacy regulations such as HIPAA and the high cost of expert annotation, poses a significant challenge for developing robust artificial intelligence (AI) diagnostic models. This study proposes a CycleGAN-based deep learning model enhanced with a histogram-guided discriminator (HisDis) to generate realistic synthetic B-mode fetal ultrasound images. A publicly available dataset from the Zenodo repository containing 1,000 grayscale fetal head images was utilized. Preprocessing included normalization, histogram equalization, and image resizing, while the architecture combined two ResNet-based generators and a dual discriminator configuration integrating PatchGAN and histogram-guided evaluation. The model was trained using standard optimization settings to ensure stable convergence. Experimental results demonstrate that the proposed HisDis module accelerates convergence by 18 epochs and reduces the Fréchet Inception Distance (FID) by 23.6 percent from 1580.72 to 1208.49 compared with the baseline CycleGAN. Statistical analysis revealed consistent pixel-intensity distributions between the original and synthetic images, with entropy from 7.16 to 7.40. At the same time, visual assessment confirmed that critical anatomical structures, including the brain midline and lateral ventricles, were well preserved. These results indicate that the CycleGAN-HisDis model produces statistically and visually realistic fetal ultrasound images suitable for medical data augmentation and AI-based diagnostic training. Furthermore, this approach holds potential to enhance diagnostic reliability and clinical education in healthcare settings with limited imaging resources. Future work will focus on clinical validation and generalization across diverse fetal ultrasound datasets.