Machine Learning Techniques for Medical Image Compression

Abstract

Medical imaging plays a vital part in modern-day healthcare practice, enabling diagnosis, treatment planning, and follow-up in the patient. Among various imaging modalities, X-ray imaging remains one of the most common due to its effectiveness, widespread availability, and diagnostic relevance. However, the continuous increase in the volume and resolution of medical images, especially X-rays, brings data storage, transmission, and processing challenges of utmost importance. These challenges normally lead to costly and complex management issues, especially in large healthcare systems. To address these challenges, both traditional compression techniques and more recent machine learning-based techniques have been explored. Traditional techniques, such as JPEG and PNG, offer basic compression but are not tailored to the specific needs of medical imaging since they are unable to preserve diagnostic quality or exploit the unique structural patterns in medical images. Yet, deep learning techniques, particularly autoencoders, convolutional neural networks (CNNs), and Variational autoencoders (VAEs), have demonstrated excellent potential in improving both image quality and compression ratios, surmounting the limitations of conventional techniques. This research investigates the potential of deep learning-based image compression models—like Autoencoders (AE), Deep Convolutional Autoencoders (DCAE), CNNs, and VAEs—for achieving better medical X-ray image compression. It also incorporates machine learning algorithms such as Principal Component Analysis (PCA) and K-means clustering for the optimization of image storage and transmission. Additionally, this study introduces the Medical X-ray Imaging Dataset (MXID), a high-quality dataset of X-ray images from AOUINET Hospital in Tebessa, Algeria, to facilitate various medical imaging tasks, including compression, classification, and machine learning applications. The study's findings demonstrate that deep learning-based models, specifically DCAEs, outperform alternative compression techniques in terms of image compression efficiency and quality retention yielding a PSNR of 46,78 dB. The ultimate objective of this research in the long term is to contribute to the development of efficient, high-quality medical image compression techniques that will be able to improve the handling of healthcare data, thereby facilitating more effective healthcare delivery through better storage and transmission.