Deep learning & hybrid model - the future of medical image watermarking

Abstract

The frequent usage of medical records in electronic form has made Medical Image Watermarking (MIW) relatively more significant than it used to be. MIW is very significant to preserve the completeness and integrity of the medical images. For the time being, with the trade-offs between visibility and robustness, there are no perfect algorithms for invisible watermarking. In many novels, Deep-Learning-Based Approach has been proposed to solve the trade-offs. In this study, multiple implementations of invisible watermarking techniques such as Deep-Learning-Based Approach and Non-Deep-Learning-Based-Approach are being compared. This comparative study measures the limitations and robustness on a dataset of breast ultrasound images. Eighteen extreme attacking methods were carried out on the encoded images, performance was then evaluated using peak signal-to-noise ratio (PSNR) and normalized cross correlation (NCC). Encoded images were then tested against a digital transmission channel to test its robustness. To conclude, The Deep-Learning-Based-Approach of RivaGAN showed the best robustness against multiple extreme attacks. The Non-Deep-Learning-Based-Approach of discrete wavelet transform – discrete cosine transform – singular value decomposition (DWT-DCT-SVD) has the best imperceptibility. Therefore, we confirm the feasibility of Deep-Learning-Based-Approach in Medical Image Watermarking, however more work is needed to be done to achieve perfect Deep-Learning-Based-Approach in terms of imperceptibility.