X-ray super-resolution stops borrowing a natural-image eye: swapping VGG perceptual loss lifts ESRGAN from 22.21 to 34.04 dB PSNR

The bottleneck in medical X-ray super-resolution is located in the loss rather than the network: one domain-aware loss can be fitted to both GAN-based and transformer-based super-resolution models, bringing the worst-performing ESRGAN back in line with the others. Trained the natural-image way with VGG perceptual loss, ESRGAN reaches only 22.21 dB PSNR and 0.5545 SSIM on chest X-rays, clearly below SRGAN, HAT and other comparison models. On PadChest chest X-rays with 500 training and 100 held-out test images, ESRGAN with USELoss rises to 34.04 dB PSNR and 0.9160 SSIM, a PSNR gain of more than 50%.

USELoss combines three constraints: a pixel loss for intensity consistency, multi-scale features from a U-Net encoder for anatomical structure, and a Sobel edge loss for boundary gradients, with a low-weight adversarial term kept only for the GAN model. Earlier perceptual losses depend on VGG networks trained on ImageNet, whose features favour texture and semantic content rather than the structural continuity and edge precision that X-rays require. The U-Net encoder has about 7.5M parameters against about 143M for VGG19; ESRGAN and HAT variants use identical loss components and weights (0.5, 1.0, 0.5, 0.005), with the adversarial term added only to ESRGAN.

Edge supervision helps both architecture families, but by different margins: the GAN model mainly uses it to suppress false edges, while the transformer model tightens boundaries that were already comparatively good. Standard ESRGAN shows the largest edge discrepancy of all models, indicating that adversarial training without domain constraints generates excessive false edges and boundary distortion. Adding USELoss cuts ESRGAN edge error by about 46%; HAT_USELoss attains the lowest edge sum (700,866) and edge mean (13.97) of all models, about 5.5% below baseline HAT.

The findings rest on a single chest X-ray dataset, PadChest, with 600 images split at the image level into 500 for training and 100 for testing; the test set was never used for model selection or tuning. Low-resolution inputs are produced by 4x bicubic downsampling, and high-resolution references are resized to 224x224 with Lanczos interpolation. Evaluation covers PSNR, SSIM, MS-SSIM, MAE, MSE, RMSE, LPIPS, VIF, FSIM, NCC, UQI, SNR and Sobel edge discrepancy, all reported as mean plus or minus standard deviation on the held-out test set. The U-Net encoder is used only during training and edge computation is light, so inference efficiency is comparable to the baseline models. Planned next steps extend the approach to more datasets and modalities such as CT and MRI, and add clinical validation through radiologist studies.

The marked degradation of standard ESRGAN on chest X-rays is attributed to VGG features favouring natural-image texture and semantics, an explanation drawn from agreement between metrics and visualisations rather than a separate ablation. HAT_USELoss has the highest SSIM, but the highest MS-SSIM belongs to baseline HAT, so cross-scale consistency does not rise with the loss swap. The method is evaluated only under bicubic degradation, does not yet cover real acquisition artefacts, and lacks clinical reading validation.