In telemedicine, digital watermarking is critical for safeguarding patient privacy by securely embedding sensitive Electronic Patient Records (EPR) directly into medical imagery. This work proposes a high-capacity, non-blind, and semi-fragile watermarking scheme leveraging a 64-bit complex-float computational pipeline and the Discrete Orthonormal Stockwell Transform (DOST) combined with Local Binary Pattern (LBP) features. To improve security, the architecture incorporates Arnold scrambling for confusion and chaotic mapping for diffusion-based encryption. In the attack-free scenario, the proposed method significantly outperformed the recent state-of-the-art (SOTA) technique across two MRI datasets. On the TCIA MRI dataset, the proposed method achieved mean values of 79.84 dB for Peak Signal-to-Noise Ratio (PSNR) and 0.9969 for Structural Similarity Index (SSIM), representing average improvements of 111.61 \(\%\) and 4.35 \(\%\) , respectively, over the SOTA baseline (37.73 dB and 0.9553). On the BMIBTD MRI dataset, the proposed method achieved mean values of 65.62 dB for PSNR and 0.9827 for SSIM, reflecting average improvements of 77.54 \(\%\) in PSNR and 11.14 \(\%\) in SSIM compared to the SOTA values (36.96 dB and 0.8842). Furthermore, the proposed method maintained a mean Normalised Correlation Coefficient (NCC) of 1, demonstrating performance parity with the SOTA technique in terms of general robustness. Extensive robustness analysis reveals a selective, tiered sensitivity profile: the scheme exhibits high resilience to stochastic channel noise (NCC \(\gtrsim\) 0.90), a characteristic semi-fragile response to Median filtering (NCC \(\approx\) 0.6–0.8), and high sensitivity to deterministic manipulations such as JPEG compression, Mean filtering, and geometric transformations. Rather than a limitation, this tiered behaviour serves as an inherent integrity-authentication mechanism, where the degradation or destruction of the mantissa-embedded watermark provides immediate detection of unauthorised spatial tampering or resampling post-acquisition. Computational time analysis demonstrates that the proposed method maintains an efficient total mean execution time of 0.2203 seconds, with separate mean processing times of 0.1293 seconds for embedding and 0.0910 seconds for extraction, ensuring rapid performance for real-time clinical applications. The capacity (8.00 bpp) and efficiency ( \(\eta\) = 0.0625) of the proposed technique were analysed and found to be high compared to recent SOTA methods. Security analysis confirms the proposed framework’s robustness against statistical, differential, and brute-force attacks, achieving superior performance metrics—including high NPCR and information entropy—compared to recent SOTA methods. Beyond MRI datasets, the system demonstrated comparable performance when validated on X-ray and ultrasound scans, confirming the framework’s cross-modality consistency. The proposed system provides a significant advancement in balancing high-fidelity diagnostic reconstruction with robust defence against cryptanalytic threats.
