Multi-layered semantic-based watermarking for medical images

Digital watermarking has drawn great attention in data integrity control and authentication lately. From the medical media perspective, one of the main applications of digital watermarking includes the embedding of patient's information into the medical media itself for the purpose of protecting the integrity of the medical media, as well as providing controls to the use of the media in the picture archiving and communication system (PACS). However, not all parts of a medical media carry the same level of importance for diagnosis and prognosis' purposes, such as digital mammograms which we are focusing on in this thesis. Furthermore, with the finite embedding capacity of the media, it is a constant challenge for one to retain the visual quality of the important regions for ensuring accurate diagnosis purposes. Therefore, there is a need for these regions of higher importance to be 'free' from visible watermarking distortion. Besides, it is also advantageous to implement some automated selection methods for these regions to eliminate the need of manual selection. In this thesis, we have implemented a reversible watermarking scheme which is capable of large capacity of data hiding, automated medical importance regions definition, running authentication to assure the integrity of the important regions on the medical images, and original image restoration. As an empirical study, we have implemented five image segmentation methods to define the segments of the image; block average, pixel classification, improve moment-preserving thresholding (IMPT), saliency map, and IMPT on saliency map. Furthermore, we conducted an analysis on both the watermarking techniques and the regions definition methods proposed in order to determine a robust combination of the above suitable for general scenarios. Besides, our scheme is capable of authentication on the protected image and verifying its integrity. The purpose for this feature is to prevent the unauthorized modification on the image which might leads to a medical misjudgment. Whenever a watermarked image is defined authentic, the scheme is capable of reconstructing the original image from the watermarked image such that medical diagnosis and prognosis process can be carried out. In addition to the above, a security analysis on the scheme is carried out to compare the proposed scheme with existing watermarking schemes with similar features and purposes. For real-time deployment, the embedding process of the proposed scheme is ported to the Compute Unified Device Architecture (CUDA) and an execution speed increase of up to 4x is observed from the empirical results.