Deep autoencoders for denoising computerized tomography (CT) images

Abstract

Understanding the complexities of human biology and making accurate medical diagnoses depend heavily on medical imaging technologies. However, a major barrier to proper analysis is the problem of noise interference in imaging. In this study, we use deep autoencoders to solve the common problem of noise reduction in computed tomography (CT) images, specifically in the 100 image Cancer Genome Atlas Lung Adenocarcinoma (TCGA-LUAD) dataset. The study measures the structural similarity index (SSIM) to evaluate the denoising performance of a sequential model using Keras. To improve the model's flexibility in real-world circumstances, Poisson noise has been introduced into the dataset. During training and validation, the model achieved significant accuracy rates of 85.63% and 88.02%, respectively, with an average SSIM score of 0.7613 for the test data. This study sheds light on the significance of deep autoencoders in advancing the domain of medical imaging, particularly in enhancing CT image reconstructions by effectively reducing noise interference while preserving crucial structural details. The findings pave the way for future refinements in deep learning methodologies tailored for medical imaging applications, offering a promising avenue toward improved diagnostic imaging in healthcare.