Enhancing Security and Robustness in Medical X-Ray Image Analysis Using CNN-GAN and Chaotic Encryption Images

Abstract

Healthcare has become increasingly fast to be digitized; this has expedited the demand of Picture Archiving and Communication Systems (PACS) as well as the Digital Imaging and Communications in Medicine (DICOM) standard of storing and transmitting of medical images. Nevertheless, their extreme sensitivity and metadata that contains patient information would make X-ray images to face the threat of unauthorized access and tampering alongside invasion of privacy. Cryptography and steganography are examples of partial protection; usually when data is decrypted or revealed, however, protection is prevented. Digital watermarking provides a solution where authentication data is embedded directly into the image itself (instead of adding an extra overhead) and yet is robust to most existing attacks but has the undesired drawbacks of being computationally intensive as well as susceptible to well-developed attacks. The present paper suggests a network-based secure medical image authentication system of X-ray images based on adversarial and deep learning networks, watermark-embedding processes based on Arnold-scrambling encryption, Finite Ridgelet Transform (FRT) and Singular Value Decomposition (SVD), optimized using firefly Algorithm (FA). Adversarial training and Generative Adversarial Networks (GANs) are also deep learning modules that increase attack resistance and reinforcement learning allows the development of adaptive watermarking. The results of experiments on the DICOM X-ray data sets indicate that the proposed approach reaches PSNR > 42 dB, SSIM > 0.9 and NC > 0.99 in no-attack conditions and NC > 0.94 in the attack with Gaussian noise, JPEG compression, median filtering and resizing. This performance shows that the framework is an encouraging solution to copyright protection, ownership verification, and safe treatment of medical images in digital healthcare systems since it has shown substantial gains against systems used in the past in terms of imperceptibility, robustness and computational simplicity.