[Paper Review] One Network to Solve All ROIs: Deep Learning CT for Any ROI using Differentiated Backprojection

Yoseob Han, Jong Chul Ye|arXiv (Cornell University)|Oct 1, 2018

Medical Imaging Techniques and ApplicationsMedicine3 references3 citations

TL;DR

This paper proposes two deep learning networks for interior computed tomography (CT) reconstruction that generalize across arbitrary region-of-interest (ROI) sizes, detector pitches, and scan angles. The first network corrects cupping artifacts in filtered back-projection (FBP) images, while the second directly inverts truncated differentiated back-projection (DBP) data using a learned finite Hilbert transform inversion, achieving superior image quality and speed over iterative methods with minimal runtime overhead.

ABSTRACT

Computed tomography for region-of-interest (ROI) reconstruction has advantages of reducing X-ray radiation dose and using a small detector. However, standard analytic reconstruction methods suffer from severe cupping artifacts, and existing model-based iterative reconstruction methods require extensive computations. Recently, we proposed a deep neural network to learn the cupping artifact, but the network is not well generalized for different ROIs due to the singularities in the corrupted images. Therefore, there is an increasing demand for a neural network that works well for any ROI sizes. In this paper, two types of neural networks are designed. The first type learns ROI size-specific cupping artifacts from the analytic reconstruction images, whereas the second type network is to learn to invert the finite Hilbert transform from the truncated differentiated backprojection (DBP) data. Their generalizability for any ROI sizes is then examined. Experimental results show that the new type of neural network significantly outperforms the existing iterative methods for any ROI size in spite of significantly reduced run-time complexity. Since the proposed method consistently surpasses existing methods for any ROIs, it can be used as a general CT reconstruction engine for many practical applications without compromising possible detector truncation.

Motivation & Objective

• To address the challenge of severe cupping artifacts in interior CT due to transverse projection truncation.
• To overcome the poor generalization of prior deep learning models trained on specific ROI sizes.
• To develop a reconstruction method that works robustly across varying detector sizes, pixel dimensions, and scan angles.
• To reduce computational complexity compared to model-based iterative reconstruction while maintaining high image quality.
• To enable practical clinical use of interior CT by creating a general-purpose reconstruction engine.

Proposed method

• Designs a Type I network that learns to remove cupping artifacts from FBP-reconstructed images, which are singular and ROI-size-specific.
• Develops a Type II network that inverts the truncated Hilbert transform from differentiated back-projection (DBP) data, using a U-Net architecture trained on DBP inputs.
• Uses data augmentation with varying detector sizes and angular starting positions to improve generalization of the Type II network.
• Applies detector and angular augmentation to reduce overfitting and enhance robustness to truncation variations.
• Employs a differentiable inversion framework that learns the null space parameters and convolutional kernels for finite Hilbert transform inversion.
• Trains both networks using the same U-Net architecture but with different input representations: FBP images for Type I and DBP data for Type II.

Experimental results

Research questions

• RQ1 Can a deep learning model generalize across different ROI sizes without retraining?
• RQ2 Does using differentiated back-projection (DBP) data as input improve generalization compared to FBP images?
• RQ3 How does data augmentation with varying detector size and scan angle affect network generalization?
• RQ4 Can a single neural network architecture effectively reconstruct high-quality images across diverse acquisition parameters?
• RQ5 Does the Type II network’s direct inversion of the finite Hilbert transform outperform post-processing of FBP artifacts?

Key findings

• The Type II network, which inverts DBP data, generalizes significantly better across ROI sizes, detector pitches, and scan angles than the Type I network.
• The Type II network achieves superior image quality compared to existing iterative reconstruction methods, with lower runtime complexity.
• Detector and angular data augmentation improved the generalization of the Type II network, though excessive detector augmentation slightly reduced performance due to averaging over diverse truncation ratios.
• The Type II network is less sensitive to sparse view and short-scan artifacts because DBP data lack the ramp filtering that induces streaking in FBP images.
• The Type II network maintains consistent performance across different acquisition scenarios, including varying detector sizes and angular starting points.
• The proposed method outperforms state-of-the-art iterative reconstruction techniques in both quantitative metrics and visual quality, with significantly faster runtime.