Hybrid intensity normalization and translation network for structure-preserving intensity normalization in medical imaging

Intensity generalization among the training and test images of one modality from different vendors is the key for better performance of deep learning models. During inference, deep learning classifiers trained on chest X-rays from one scanner often fail when applied to images from a different scanner due to contrast variations. With Shenzhen–Montgomery tuberculosis benchmark, the drop can be as large as 0.95 to 0.73 in AUC. We present HINT-Net, a single-generator framework for unpaired, multi-domain intensity normalization of medical images. The architecture builds on a U-Net backbone and introduces two mechanisms: (i) Feature-wise Linear Modulation (FiLM) layers that conditions every encoder and decoder stage on a target-domain vector so that one model can serve any number of domains without retraining and (ii) an Additive Spatial Gate (ASG) at the bottleneck that learns a pixel-level intensity correction map to suppress scanner-specific artifacts while leaving anatomy intact. Training combines a hybrid loss of adversarial, Patch-based Noise Contrastive Estimation (PatchNCE), Structural similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), and gradient terms. On paired synthetic intensity domains from NIH ChestX-ray14, HINT-Net reaches SSIM above 0.98, LPIPS below 0.008, and Peak Signal-to-noise Ratio (PSNR) above 34 dB-outperforming CycleGAN by 3.3% in SSIM with 86% lower LPIPS. On the unpaired Shenzhen–Montgomery benchmark it achieves SSIM of 0.78 (versus 0.74 for CycleGAN) and raises area under the curve of tuberculosis detection from 0.64 to 0.91 with AlexNet and from 0.78 to 0.97 with ResNet-18.