Evaluating the Performance of Compressive Sensing and Deep Learning Techniques for Medical Imaging in Healthcare Applications

Evaluating the Performance of Compressive Sensing and  Deep Learning Techniques for Medical Imaging in Healthcare Applications

Umme Saniya 1, Yashass G2, Soujanya 3, Chandini R 4. Dr. R. SEKAR M.E.,PH.D

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

PRESIDENCY UNIVERSITY

ABSTRACT:

In recent years, the use of deep learning- based medical image fusion (DLMIF) has become a common practice for the reliable detection of diseases, as it allows the combination of information from multiple medical imaging modalities [1], [2]. The performance of DLMIF heavily relies on the effective selection of features for calculating fusion weights [3]. This study investigates the efficacy of convolutional neural network (CNN) features in the context of DLMIF by utilizing two input medical images and generating a fused image using various conventional techniques [4]. Due to the absence of ground-truth images for training end-to-end CNNs, pre-trained networks from other tasks are employed to extract relevant features [5]. The choice of CNN network and the selection of convolutional layers (CL) are systematically examined to assess their impact on the fusion process [6]. Furthermore, consistency maps and local visibility are computed using the extracted CNN feature maps to determine the appropriate fusion weight map for DLMIF [7]. The results demonstrate that the proposed method outperforms conventional techniques in terms of several quantitative metrics and produces superior DLMIF outputs, offering enhanced medical images that are highly suitable for medical diagnosis [8].

Parallel to advancements in medical image fusion, the increasing use of images across various sectors, including online social networks, government agencies, law enforcement, educational institutions, and

private companies, has driven the demand for efficient image storage solutions [9]. As these images are stored in vast databases, image compression techniques play a critical role in reducing storage requirements and optimizing data transfer [10]. Image compression aims to represent significant image information in a compact form while removing redundant or insignificant data [11]. The rapid growth of data has highlighted the importance of efficient image compression, especially in the face of the challenges posed by complex, unknown correlations between pixels in an image [12]. The task of finding and recovering well-compressed representations is intricate, and designing networks that can recover images successfully—either losslessly or lossy—remains a challenging task [13]. Deep learning techniques, particularly autoencoders, have gained attention as effective tools for image compression [14].

This article provides an overview of the most common image compression techniques, focusing on the role of autoencoders in deep learning-based compression, and evaluates key performance metrics such as SSIM, MS- SSIM, and PSNR to assess the effectiveness of these methods in maintaining image quality during compression [15]. By integrating advancements in both DLMIF and image compression, this study emphasizes the potential of deep learning techniques to improve medical image analysis and data storage solutions across various fields [16].