Stable training and confidence‑calibrated generation of urinary stone CT images via physics‑constrained GANs under small‑sample conditions

In recent years, artificial intelligence has become integral to modern medicine. Deep learning algorithms are widely used to assist clinical diagnosis, and medical image synthesis has become increasingly mature. However, in small-sample scenarios, synthetic images often overlook infrequent but clinically valuable lesion features. To address the challenges of training instability, insufficient authenticity, and limited clinical usability, we propose a two-stage framework based on Generative Adversarial Networks (GANs). In the first stage, we establish a stable training and physical prior constraint framework for small-sample CT images. Specifically, we incorporate an established asymmetric normalization strategy into the WGAN-GP framework to balance generator convergence and the discriminator's Lipschitz constraint, and we embed a physical noise simulation layer at the discriminator's front end, combined with partial volume effect gradient regularization, to enhance the realism of synthesized images. In the second stage, we employ MC Dropout and deep ensemble methods to quantify uncertainty, generating uncertainty heatmaps and physically consistent confidence maps based on the Poisson-Gaussian noise model. This assists physicians in locating abnormal areas and reducing misdiagnosis risks. Through joint optimization of the generation framework, uncertainty quantification, and confidence image output, the proposed framework effectively resolves issues of mode collapse, low confidence, and poor clinical usability under small-sample conditions. The proposed framework achieves a competitive FID of 128.0, an Expected Calibration Error (ECE) of 0.072, and a high-risk recall rate of 88% for stone boundary regions, demonstrating its potential for clinically reliable synthetic CT generation.