A Fast, Scalable, and Robust Deep Learning-based Iterative Reconstruction Framework for Accelerated Industrial Cone-beam X-ray Computed Tomography

TL;DR

This work tackles the challenge of fast, high-quality 3D reconstructions in industrial cone-beam XCT for large volumes by marrying a pre-trained artifact-removal CNN with a model-based HQS inversion. The algorithm alternates CNN-based proximal updates with a CG data-consistency step and introduces an automatic, per-iteration regularization parameter selection guided by BRISQUE on center slices, significantly reducing outer iterations. It demonstrates strong generalization to out-of-distribution scans across varying voltage, integration time, and sparsity, while offering substantial runtime gains over MBIR. The approach delivers robust artifact suppression and pore preservation in high-density materials, enabling near in-line operation for high-throughput industrial XCT workflows, with memory-efficient training and scalable inference on large 3D volumes.

Abstract

Cone-beam X-ray Computed Tomography (XCT) with large detectors and corresponding large-scale 3D reconstruction plays a pivotal role in micron-scale characterization of materials and parts across various industries. In this work, we present a novel deep neural network-based iterative algorithm that integrates an artifact reduction-trained CNN as a prior model with automated regularization parameter selection, tailored for large-scale industrial cone-beam XCT data. Our method achieves high-quality 3D reconstructions even for extremely dense thick metal parts - which traditionally pose challenges to industrial CT images - in just a few iterations. Furthermore, we show the generalizability of our approach to out-of-distribution scans obtained under diverse scanning conditions. Our method effectively handles significant noise and streak artifacts, surpassing state-of-the-art supervised learning methods trained on the same data.

Figures (9)

• Figure 1: Desired properties of computed tomography reconstruction algorithms required for high-quality imaging of large 3D volumes encountered in industrial CT settings. Existing approaches in the literature satisfy two of the three properties.
• Figure 2: Schematic of a cone-beam X-ray computed tomography with common nomenclature.
• Figure 3: Illustration of the proposed algorithm. Our method alternates between an artifact removal CNN and a few iterations of a conjugate gradient algorithm that enforces data-consistency while ensuring the result is close to the output of the CNN. Additionally in each outer iteration, the regularization parameter for the CG stage is adjusted based on as assessment of image quality. This architecture ensures a high-quality reconstruction with few outer iterations while being more generalizable compared to vanilla single-step CNNs.
• Figure 4: Reconstruction performance evaluation of the proposed method in terms of PSNR (in dB) and SSIM on sample ST at 160 kV voltage and 0.6 seconds integration time. Plots (a) and (b) are denoting performance using either fixed or adaptive regularization parameter selection strategy while (c) and (d) represent performance for different iterations of the conjugate gradient algorithm. The curves in (c) and (d) are for the CG in four different outer loops, $k = 1, 2, 3, 4$ respectively.
• Figure 5: Illustration of the performance on ST sample from Table \ref{['tab:ct_data']} scanned under different integration times of 0.6 seconds, 1.8 seconds and 3.6 seconds (scan time of 10 minutes, 30 minutes, 60 minutes respectively). The peak source voltage is 200 kV for each case. This is an out-of-distribution data since the training data for DLMBIR and the proposed were from an 8 second integration time. BHCN-MBIR on data from 200 kV, 3.6 seconds integration time (60 minutes scan time) is used as the reference. The yellow arrows point out the ring artifacts that are more prominent in BHCN-FDK and BHCN-DLMBIR reconstructions from lower integration times. The proposed method gets rid of these artifacts for all the integration times shown here. The green arrow shows the pores that have been well preserved by both BHCN-DLMBIR and the proposed method.