Privacy Preserving Medical Image Classification and Steganography Using Deep Learning Architecture: a Pipeline

Secure transfer of medical data over unsecured networks is a crucial concern in telemedicine, as it involves the transmission of sensitive patient information that must be protected from unauthorized access. Generative adversarial networks help in hiding arbitrary data in images while optimizing the quality of the images. In this study, we introduce a novel pipeline that classifies medical images, integrates the diagnosis with patient information, embeds the concatenated information in patient image data, and ensures robust sharing capabilities. The pipeline comprises of image classification and image steganography. Image classification employs fine-tuned state-of-the-art (SOTA) convolutional models. Steganography makes use of encoder-decoder architecture to embed information on the image. These methods were tested on two medical datasets (X-ray and CT scans). Several models for classification were tested, of which fine-tuned Densenet121 achieved an accuracy and F1 score of 98.87 % and 98.8 % on X-ray respectively, and 98.1 % and 97.8 % on CT scans respectively. Similarly, for steganography our approach achieves payloads of 2.68 bits per pixel for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$X$</tex> ray image and 2.56 bits per pixel for chest CT scan, Peak signal to noise ratio (PSNR) of 51.15 db for X-ray and 53.3 db for CT scan, and structural similarity index measure (SSIM) of 0.99 and 0.99 for X-ray and CT scan dataset respectively.