Spatially-Adaptive Reconstruction in Computed Tomography using Neural Networks

Joseph Shtok, Michael Zibulevsky, Michael Elad

Introduction

Computed Tomography (CT) imaging is a cornerstone of modern diagnostics, yet it faces a persistent challenge: the trade-off between radiation dose and image quality. Reducing X-ray exposure—essential for patient safety—introduces stochastic noise and "streak" artifacts. Traditional reconstruction methods, such as Filtered Back-Projection (FBP), require a manual choice of filtering parameters. A sharp filter preserves detail but amplifies noise, while a smooth filter reduces noise at the cost of blurring critical diagnostic features. This paper, published in 2013, proposes a spatially-adaptive solution: using an Artificial Neural Network (ANN) to intelligently fuse multiple reconstructions of the same scan, optimizing for both clarity and noise suppression.

Methods: Locally-Adaptive Fusion

The core innovation is a supervised learning framework that treats different reconstruction parameters as complementary information sources.

Multi-Version Generation and Feature Extraction

For a single CT scan, the authors generate $K$ different reconstructions. In the case of FBP, this involves varying the filter's cut-off frequency. For iterative methods like Penalized Weighted Least Squares (PWLS), snapshots are taken at different iteration counts. For every pixel $q$, a small neighborhood (a disk of radius 3 pixels) is extracted from all $K$ versions to serve as input features.

Neural Network Architecture

These local neighborhoods are fed into a feed-forward ANN. The network is trained to map these $K$ noisy versions to a single, high-quality output patch. To ensure the network prioritizes diagnostic relevance, a weighted loss function is employed during training, ignoring "air" regions and focusing on informative edges and tissue textures.

Architecture of the feed-forward Artificial Neural Network used for pixel-wise fusion.

To prevent boundary artifacts between processed regions, the network outputs small patches rather than individual pixels. Overlapping patches are then averaged to produce the final, spatially-consistent image.

Datasets and Training

The model was trained using clinical CT data from the Visible Human Project. Specifically, 12 axial slices of male thigh sections provided 30,000 local patches for supervised learning. High-dose, high-quality scans served as the ground truth. To evaluate robustness, the method was tested on slices from different anatomical regions, including the head and abdomen, located significantly far from the training data.

Key Results and Insights

The proposed ANN fusion acts as a non-linear, spatially-varying filter. It learns to "trust" smooth reconstructions in homogeneous areas to eliminate noise, while relying on sharper versions near high-contrast interfaces like bone or blood vessels.

Quantitative and Qualitative Performance

The method demonstrated significant improvements across multiple metrics:

  • Signal-to-Noise Ratio (SNR): For FBP, the fused result achieved a 1.5 dB to 1.8 dB improvement over the best single-parameter reconstruction.
  • Structural Similarity (SSIM): The ANN-fused images consistently scored higher on SSIM, which aligns more closely with human visual perception than standard mean squared error.
  • Visual Quality: Qualitatively, the images showed a dramatic reduction in streaking artifacts while maintaining sharp bone-flesh interfaces.
Comparison between a standard reconstruction (left) and the ANN-fused result (right).

One of the most compelling findings was the "universality" of the learned features. Despite being trained only on thigh scans, the network performed remarkably well on head scans, suggesting that local geometric features like edges and noise patterns are consistent across human anatomy.

Conclusion and Future Work

This early application of machine learning to CT reconstruction proved that spatially-adaptive fusion can bypass the fundamental trade-offs of classical algorithms. While the 2013 study utilized a relatively shallow architecture, it laid the groundwork for modern deep learning approaches in medical imaging. Future work could involve automating the selection of input parameters ($K$) and exploring deeper convolutional architectures to further enhance feature extraction and reconstruction speed.

Summary updated

Abstract

We propose a supervised machine learning approach for boosting existing signal and image recovery methods and demonstrate its efficacy on example of image reconstruction in computed tomography. Our technique is based on a local nonlinear fusion of several image estimates, all obtained by applying a chosen reconstruction algorithm with different values of its control parameters. Usually such output images have different bias/variance trade-off. The fusion of the images is performed by feed-forward neural network trained on a set of known examples. Numerical experiments show an improvement in reconstruction quality relatively to existing direct and iterative reconstruction methods.
BibTeX
@article{shtok2015spatially,
  author = {Joseph Shtok and Michael Zibulevsky and Michael Elad},
  title = {{Spatially-Adaptive Reconstruction in Computed Tomography using Neural Networks}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2015},
  doi = {10.1109/tmi.2015.2401131},
}
Figure