Explainable AI-Driven Robustness and Clinical Safety Evaluation of Medical Image Watermarking Schemes

Abstract

Medical image watermarking plays a pivotal role in safeguarding clinical imaging processes in Picture Archiving and Communication Systems (PACS) and teleradiology applications. Existing evaluation frameworks rely solely on global signal metrics, such as the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), which measure average distortion and cannot determine whether the distortion falls within diagnostically critical image regions. This study proposes an explainable artificial intelligence (XAI)-driven framework for evaluating the clinical safety of medical image watermarking schemes across three imaging modalities. Three DenseNet-121 diagnostic models were fine-tuned on publicly available CT, chest X-ray (CXR), and MRI datasets, achieving validation accuracies of 97.5%, 77.5%, and 90.8%, respectively. Grad-CAM saliency maps derived from these models served as the basis for three proposed clinical safety metrics: Saliency Overlap Score (SOS) (Dice/IoU), Diagnostic Probability Shift (ΔP), and Region-specific PSNR/SSIM. A novel XAI-guided watermarking scheme was proposed that restricts embedding to low-saliency background pixels. A total of 7,200 evaluations were conducted across four watermarking schemes, six attack conditions, and three imaging modalities. The XAI-guided scheme achieved the highest SOS Dice on all three modalities (CT: 0.896±0.088, CXR: 0.855±0.168, MRI: 0.935±0.105) and the lowest ΔP (CT: 0.008±0.012, CXR: 0.011±0.009, MRI: 0.004±0.007). An early training HiDDeN-based distortion-injection baseline, despite achieving an ROI-PSNR of 52–71 dB, satisfied the proposed ΔP < 0.02 reference criterion in only 31–68% of images per modality, suggesting that a high global PSNR does not reliably indicate clinical safety for learned-perturbation watermarking. The proposed framework demonstrates that saliency-aware evaluation provides clinically relevant insights beyond conventional signal fidelity metrics and offers a practical methodology for assessing the clinical safety of medical image watermarking schemes. The findings further suggest that the preservation of diagnostic saliency and model confidence should be considered alongside PSNR, SSIM, and BER when evaluating the watermarking performance in clinical imaging environments.