Study of Analytic vs Algebraic Image Tomography Reconstruction (Conventional and Deep Learning)

Abstract

Image reconstruction in tomography is a key challenge in medical imaging, especially in complex clinical situations involving incomplete data, noise, or low-dose acquisition. This work focuses on the study and implementation of the most efficient tomography reconstruction techniques based on a parallel beam model, ranging from conventional methods to modern approaches powered by artificial intelligence. Our contribution is the enhancement of the reconstructed images especially in noisy data, through the use of convolutional neural networks (CNN). Two main categories were explored: analytical methods, represented by Back Projection (BP) and Filtered Back Projection (FBP), and iterative methods, including the Simultaneous Algebraic Reconstruction Technique (SART) and the Maximum Likelihood Expectation Maximization (MLEM) algorithm. To enhance reconstruction quality, these methods were extended with Convolutional Neural Networks (CNNs). Four hybrid architectures were developed: FBP + DnCNN, Learnable FBP, MLEM + DnCNN, and Learnable MLEM. Experiments were conducted on both synthetic images and real medical scans, and evaluated in terms of quantitative metrics (MSE, PSNR, SSIM, 𝐷𝑓,𝐷𝑝) in addition to visual quality. The obtained results demonstrated a significant improvement of hybrid models over conventional ones, particularly the Learnable MLEM, which yielded the most accurate reconstructions in terms of metrics and visual quality. Learnable schemes in both MLEM and FBP are more efficient than the combination with DnCNN. This study highlights the relevance of CNN-based hybrid approaches in parallel-beam tomography, which in turn allow precision, robustness, and clinical useful reconstructions.