Map-Assisted Lung-Aware Carrier Selection for Chest-Radiograph Watermarking: A Fixed-Configuration Cross-Dataset Evaluation

Watermarking chest radiographs requires limiting changes within lung fields while retaining payload recovery. We evaluated a map-assisted, host-image-independent transform-domain framework using three paired carrier-selection strategies: Global, reference-mask-guided GT-RONI, and U-Net-guided AI-RONI. A 112-bit message, rate-1/3 channel code, embedding strengths, and decoder were fixed on the Japanese Society of Radiological Technology (JSRT) dataset and transferred without retuning to 138 Montgomery radiographs. Relative to Global, AI-RONI reduced reference lung mean squared error by 51.9% internally and 45.7% externally. No-attack exact recovery was 20/20 internally (95% Wilson interval, 83.9–100.0%) and 84/138 externally (60.9%); GT-RONI achieved 60.1% externally. Controlled replacements characterized the balance between fidelity and recovery: the coefficient-pair alternative to singular value decomposition achieved 100.0% external no-attack recovery but zero recovery under the three tested processing operations, while omission of the discrete cosine Stockwell transform produced little change. Repeated Gaussian-noise and carrier-map-corruption tests further characterized decoder sensitivity. The framework provides a controlled basis for assessing anatomical exposure and payload transfer jointly across image sources.