[Paper Review] Artifact Reduction in Undersampled 3D Cone-Beam CTs using a Hybrid 2D-3D CNN Framework

Johannes Thalhammer, Tina Dorosti+4 more|arXiv (Cornell University)|Feb 9, 2026

Medical Imaging Techniques and Applications0 citations

TL;DR

This paper presents a hybrid 2D-3D CNN framework that uses a 2D U-Net to extract slice-wise features and a 3D decoder to fuse them, reducing artifacts in undersampled 3D cone-beam CT while maintaining low computational overhead.

ABSTRACT

Undersampled CT volumes minimize acquisition time and radiation exposure but introduce artifacts degrading image quality and diagnostic utility. Reducing these artifacts is critical for high-quality imaging. We propose a computationally efficient hybrid deep-learning framework that combines the strengths of 2D and 3D models. First, a 2D U-Net operates on individual slices of undersampled CT volumes to extract feature maps. These slice-wise feature maps are then stacked across the volume and used as input to a 3D decoder, which utilizes contextual information across slices to predict an artifact-free 3D CT volume. The proposed two-stage approach balances the computational efficiency of 2D processing with the volumetric consistency provided by 3D modeling. The results show substantial improvements in inter-slice consistency in coronal and sagittal direction with low computational overhead. This hybrid framework presents a robust and efficient solution for high-quality 3D CT image post-processing. The code of this project can be found on github: https://github.com/J-3TO/2D-3DCNN_sparseview/.

Motivation & Objective

• Motivate reducing artifacts in undersampled CT volumes to enable faster acquisition and lower radiation dose.
• Propose a hybrid architecture combining 2D slice-wise feature extraction with 3D volumetric decoding.
• Evaluate artifact reduction performance against sparse-view data using PSNR and SSIM metrics.

Proposed method

• Train a 2D U-Net on 512x512 axial slices to remove artifacts from sparse-view inputs.
• Extract and stack 2D feature maps from a volume to form a 3D feature representation.
• Use a 3D decoder with 3x3x3 convolutions to reconstruct an artifact-reduced 3D volume from the stacked features and sparse input.
• Compare 2D-only and 3D decoder post-processing in terms of PSNR, SSIM, and computational time.

Fig. 1: Overview of the 2D-3D training pipeline. A) First, a 2D U-Net is trained on 2D axial slices to remove artifacts. B) The encoder of the trained U-Net is used to extract features at different levels of an entire CT volume by looping through $N$ axial slices. Subsequently, the extracted feature

Experimental results

Research questions

• RQ1 Can a hybrid 2D-3D CNN framework improve artifact reduction for undersampled cone-beam CTs compared to 2D or 3D models alone?
• RQ2 Does stacking 2D features into a 3D representation provide volumetric consistency improvements without prohibitive computation?
• RQ3 What are the quantitative gains in PSNR and SSIM and the runtime trade-offs for the proposed approach?

Key findings

• The 2D U-Net and the 3D decoder both substantially improve PSNR and SSIM over sparse-view reconstructions.
• In axial views, both methods yield visually comparable artifact reduction; in coronal and sagittal views, the 3D decoder offers superior inter-slice consistency.
• PSNR on test data: sparse-view 24.810 ± 0.522; 2D U-Net 39.289 ± 1.295; 3D decoder 38.086 ± 1.188.
• SSIM on test data: sparse-view 0.637 ± 0.015; 2D U-Net 0.949 ± 0.015; 3D decoder 0.938 ± 0.016.
• Runtime per volume: 2.408 ± 0.651 seconds for 2D U-Net post-processing; 20.276 ± 5.768 seconds for the 3D decoder including feature extraction (on NVIDIA A100).

Fig. 2: An axial, coronal, and sagittal slice of the full-view reconstruction from the test split (first row), 128-view reconstruction (second row) with artifact reduction by the 2D U-Net (third row) and by the Hybrid 2D-3D CNN framework (last row), respectively. Images are in the mediastinum window