ENHANCED DATA HIDING BY SEGMENTATION FOR MEDICAL IMAGES

Abstract

Due to the large amount of data flowing across internet, there is a need to safeguard the transmission of sensitive data, specifically in the field of utmost security and privacy. Data hiding significantly contributes to protect the sensitive information by embedding the data inside a cover media in such a way that is undetectable to human eyes. Due to the low contrast of medical images, it often suffers from distortion after embedding the data. In order to resolve this, most of the researchers incorporate image segmentation to segment the medical image into Region OF Interest (ROI) and NROI (Non-Region Of Interest). This helps in performing the selective embedding based on the features of different regions. However, existing methods suffer from accurately separating the ROI from the rest of the medical image, which can impact the data hiding performance. Therefore, in order to solve this problem, this report introduces an improved UNet 3+ architecture, which is based on UNet framework. It overcomes the discrepancies faced by the base UNet model with the ground truth segmented mask. This enhanced network integrates residual blocks at both encoder and decoder to handle gradient degradation problem in deep neural network, incorporates attention module to focus on relevant features and reduces redundant skip connections to reduce computational complexity. The improved model has been evaluated on publicly available PH2 skin lesion dataset and achieved an accuracy of 96.79 %, dice score of 93.72 % and jaccard index of 96.97 % on test data. The model surpasses other state-of-the-art UNet models for skin lesion segmentation. After segmenting the ROI, a conventional data hiding method has been used for selective embedding based on the local complexity of the pixels. This local complexity helps in embedding only in smoother region of the ROI to reduce the distortion. The embedding capacity is kept low to avoid affecting the most critical region but the visual quality of stego image has been enhanced with PSNR of 33.08 dB and SSIM value of 0.9893.