Evaluation of DeepMedic Neural Network for a Region of Interest Eextraction in Medical Image Watermarking
Anna Yu. Denisova · 2023
Medical image watermarking prevents data modification and falsification by third parties. For the reason of safety, the key part of the image used for diagnosis should be protected by watermarks that are stable to different attacks. However, watermarks produce little distortions that might influence automatic diagnostic procedures. The best way to overcome this is to divide the image into a region of interest (ROI) and a region of noninterest (RONI) where ROI is protected by a fragile watermark and RONI is served for storing correction coefficients required to restore ROI data precisely. To make the watermarking system fully automatic, ROI detection methods are needed. In this paper, it is proposed to use a DeepMedic neural network for ROI extraction against using state-of-the-art manual ROI extraction and binary thresholding. The manual ROI extraction is difficult because of additional human labor while thresholding usually leads to homogenous RONI where the correction coefficients cannot be effectively stored. In this paper, an evaluation of DeepMedic parameters is provided in the case of the ROI extraction for magnetic resonance imaging (MRI) data. The paper demonstrates that selected parameters give better semantic segmentation results than the original ones reported by the DeepMedic network authors.