Deep Guess acceleration for explainable image reconstruction in sparse-view CT

TL;DR

This work tackles sparse-view CT reconstruction by coupling a deep-learning initial guess with a non-convex TpV-regularized MBIR solver. The Deep Guess step provides a fast, data-driven starting point, which together with a few Chambolle-Pock iterations yields accurate, interpretable reconstructions. Across real and synthetic datasets, the approach accelerates convergence, reduces susceptibility to poor local minima, and remains robust to noise, offering 15–25× faster performance while maintaining high image quality. The framework accommodates ground-truth-free training and flexible component choices, enabling practical deployment in clinical imaging alongside principled optimization.

Abstract

Sparse-view Computed Tomography (CT) is an emerging protocol designed to reduce X-ray dose radiation in medical imaging. Traditional Filtered Back Projection algorithm reconstructions suffer from severe artifacts due to sparse data. In contrast, Model-Based Iterative Reconstruction (MBIR) algorithms, though better at mitigating noise through regularization, are too computationally costly for clinical use. This paper introduces a novel technique, denoted as the Deep Guess acceleration scheme, using a trained neural network both to quicken the regularized MBIR and to enhance the reconstruction accuracy. We integrate state-of-the-art deep learning tools to initialize a clever starting guess for a proximal algorithm solving a non-convex model and thus computing an interpretable solution image in a few iterations. Experimental results on real CT images demonstrate the Deep Guess effectiveness in (very) sparse tomographic protocols, where it overcomes its mere variational counterpart and many data-driven approaches at the state of the art. We also consider a ground truth-free implementation and test the robustness of the proposed framework to noise.

Figures (8)

• Figure 1: Graphical representation of the proposed data-driven approach for solving the non-convex imaging problem.
• Figure 2: The Deep Guess block. It can be computed by following the red, orange, or green path, all exploiting a neural network. The input to the network is a coarse reconstructed image, obtained either through FBP or after $k$ iterations of an MBIR solver. The target images are either the ground truth or those obtained when the MBIR method has converged.
• Figure 3: Ground truth images from the test set of Mayo Clinic (left) and COULE (right), with two zooms-in on the regions depicted with red rectangles on the corresponding whole images.
• Figure 4: Results on the Mayo test image shown in Figure \ref{['fig:gt']}, computed using the TV prior (first column), the T pV prior (second column), and the W regularizer (third column), in case of $\mathcal{G}_{180,60}$ sparse protocol.
• Figure 5: Results on the Mayo Clinic test image shown in Figure \ref{['fig:gt']}, computed using the CP algorithm (first column) and the proposed FBP-LPP and TV-RISING Deep Guess approaches (second and third columns), within the $\mathcal{G}_{360,360}$ tomographic protocol and setting $p=0.5$ for the final CP execution.