PSWGAN-GP and MedicalDISTS synergistic optimization framework for few-shot fetal ultrasound image augmentation

The scarcity of fetal ultrasound image samples significantly constrains the performance of fetal anomaly screening models. To overcome the limitations of conventional augmentation techniques-including speckle noise-induced blurring, structural distortions, training instability, and the absence of effective sample filtering-this paper proposes a data augmentation framework synergistically driven by perceptually stabilized Wasserstein generative adversarial network with gradient penalty and MedicalDISTS (medical deep image structure and texture similarity). The framework incorporates a spectral normalized attention residual network to enhance organ contour clarity, a multi-scale deep perceptual feature pyramid to correct morphological distortions, and a multimodal stability-enhanced training mechanism to suppress mode collapse. The MedicalDISTS module further employs an edge-aware Sobel operator, a ResNet50-convolutional block attention module feature extractor, and adaptive thresholding to identify and filter structurally anomalous samples. Experiments on fetal cranial ultrasound data show that the proposed method reduces the Fréchet inception distance to 50.02 with a 48.73% decrease in model parameters. After screening, the dataset size is reduced by 43.53%, while downstream task performance is significantly improved, achieving a mAP@0.5:0.95 of 0.927 in detection and an Intersection over Union of 0.972 in segmentation. This demonstrates effective quality-quantity synergy, offering a reliable solution for small-sample fetal ultrasound image enhancement.