Image Super Resolution for Medical Images

Medical imaging plays a critical role in diagnosing diseases via modalities such as chest X- rays, but very low-resolution images hinder analysis. Traditional upsampling methods like bicubic interpolation fail to recover fine details, leading to blurred features. In this work, we develop Generative Adversarial Network GAN based super-resolution models Super- Resolution Generative Adversarial Network SRGAN) and Swift- SRGAN) to upscale chest Xray images while preserving diagnostic structures. Our proposed pipeline first normalizes and augments the National Institutes of Health NIH Chest X- ray data, then applies bilateral filtering for edge- preserving denoising. The denoised images are passed to the super-resolution generator. SRGAN uses deep generator–discriminator networks with adversarial and perceptual losses to enhance fine details, whereas Swift- SRGAN employs a more lightweight architecture to reduce computation while maintaining high fidelity. We train both models on a Graphics Processing Unit GPU (using PyTorch/TensorFlow) with tuned hyperparameters (e.g. Adam optimizer, learning rate 1e-4, batch size [16]. Quantitatively, our method achieves 13 d B higher Peak Signal- to-Noise Ratio PSNR 47.5 0 d B v s 34 .5 0 d B than bicubic upsampling and Structural Similarity Index Measure SSIM > 0. 98 on test X- rays, representing a 1 5 % improvement in SSIM over traditional methods. A Streamlit-based web interface allows clinicians to upload and interactively compare original versus enhanced images. Experimental results show visibly sharper reconstructions that retain critical anatomical structures, indicating feasibility for clinical use.