Medical image noise severely hinders accurate clinical diagnosis, yet traditional deep learning methods often cause over-smoothing of high-frequency textures and are too heavy for edge deployment. To address this, we propose a lightweight GAN with a U-Net-based generator that integrates a Texture Enhancement Module (TEM) with adaptive attention and residual fallback to dynamically preserve fine-grained details. A joint loss function combining L1, VGG perceptual, and PatchGAN adversarial losses shifts the optimization from pixel-wise error to perceptual similarity. Evaluated on a physically simulated blood-water endoscopic dataset, our method achieves the best LPIPS (0.2940), LBP (0.8598), and UIQM (9.6457), providing an efficient high-fidelity restoration solution for medical edge computing.