Efficient Reconstruction and Secure Transmission of Medical Images For Telemedicine Applications

Abstract

Telemedicine has become a crucial solution for delivering healthcare services remotely, particularly in remote areas and during emergencies. It enhances communication between patients and healthcare providers by enabling remote monitoring, medical imaging, and data sharing. By integrating smart healthcare devices, telemedicine systems can transmit real-time patient data directly to physicians, supporting timely consultations and interventions. Store-and-forward telemedicine services allow for the collection of medical data—including images, pathology reports, and patient health information—at one location, where it is securely stored and later transmitted to healthcare providers at a different location for review and consultation. This thesis presents a comprehensive approach for the reconstruction and secure transmission of low-dose computed tomography (LDCT) images using store and forward telemedicine services. In the first work we presented a comprehensive review based on computational medical image reconstruction techniques. This study aims to discuss some of the significant contributions of data-driven techniques to solve the inverse problems in medical image reconstruction (MIR). In this study, we have comprehensively studied various reconstruction techniques based on machine learning and deep learning. This work provides a detailed survey of MIR, which includes the traditional reconstruction algorithm, machine learning, and deep learning-based approaches such as GAN, autoencoder, RNN, U-net, etc., to solve inverse problems, evaluation metrics, and openly available codes used in the literature. The second work introduces unsupervised image reconstruction techniques that combine iterative reconstruction techniques, i.e., analytical and statistical methods with neural networks. In the first part, the study integrates the Maximum Likelihood Expectation Maximization (MLEM) algorithm with unsupervised deep convolutional neural network (DCNN) priors. The second approach presents a novel unsupervised CT reconstruction method that leverages an Attention-Enhanced Deep Image Prior (AE-DIP) in fusion with the Simultaneous Algebraic Reconstruction Technique (SART). The third approach introduces Loss-construct Unsupervised Network Adjustment (LUNA) for low-dose CT reconstruction. It combines SART with weighted total variation (WTV) regularization within a Deep CNN, optimized using the ADMM framework for balanced and accurate results. Multiple loss functions—perceptual, SSIM, WL2, WTV, and sinogram loss—guide the network updates, addressing data limitations in deep learning. This robust, unsupervised method effectively enhances low-dose CT image reconstruction. v The fourth approach is a pixel-based blind self-embedding fragile watermarking technique tailored for e-healthcare applications, specifically designed for store-and-forward telemedicine services. Authentication bits are generated through Chaotic Coordinate Mapping, which involves a series of logical and cyclic operations within the spatial features of the image. The Linear Feedback Shift Register (LFSR) is employed to generate both a random matrix and a cryptic key. Ensuring perceptual quality after the watermark insertion is necessary for CT images, specifically for low-dose images. Keywords: Store and forward telemedicine services, inverse problems, low CT image reconstruction, Iterative image reconstruction methods, Deep image Prior, ADMM optimizations, loss functions, CT image authentication, Blind watermarking, self-embedding fragile watermarking.