204. Preprocessing and Hybrid Strategies for Noise Characterization and Image Enhancement in Medical Tomography

Abstract:

Noise and reconstruction artifacts limit the diagnostic value of CT and MRI images and affect the reliability of automated analysis such as segmentation and AI-based tools. This talk presents the results of my doctoral research on preprocessing methods that reduce noise while preserving the anatomical structures needed for interpretation. I begin by showing that, in the CT data studied, noise is not simply Gaussian: statistical and deep-learning analyses indicate Poisson noise dominates in sharp and medium reconstruction kernels. I then introduce quaternion bilateral filtering, which converts sharp-kernel CT images into soft-kernel-like images without access to raw projection data. Next, I combine variational mode decomposition with quaternion filtering to denoise selectively. Finally, I present a hybrid framework based on decomposition, Curvelet thresholding and classical filters, evaluated on CT, MRI and phantom data and compared with BM3D and DnCNN. I will discuss the results, the limitations and directions for future work.

About the author:

Mahmoud Nasr is a doctoral student in the Computational Neuroscience team at Sano Centre for Computational Medicine and a PhD candidate in Biomedical Engineering at AGH University of Krakow. With a background in electrical engineering, specializing in electronics and communication systems, his research focuses on noise characterization and denoising of CT and MRI images using quaternion, decomposition-based and multi-scale methods.