[Paper Review] Deriving Neural Network Architectures using Precision Learning: Parallel-to-fan beam Conversion

Christopher Syben, Bernhard Stimpel+4 more|arXiv (Cornell University)|Jul 9, 2018

Advanced MRI Techniques and Applications7 references4 citations

TL;DR

This paper proposes a precision learning-based neural network for parallel-to-fan beam conversion in hybrid MRI/X-ray imaging, replacing interpolation-heavy rebinning with a differentiable, convolution-based filter learning approach. The method achieves sharper images by avoiding resolution loss, learning projection-dependent and independent filters from synthetic data with minimal trainable parameters.

ABSTRACT

In this paper, we derive a neural network architecture based on an analytical formulation of the parallel-to-fan beam conversion problem following the concept of precision learning. The network allows to learn the unknown operators in this conversion in a data-driven manner avoiding interpolation and potential loss of resolution. Integration of known operators results in a small number of trainable parameters that can be estimated from synthetic data only. The concept is evaluated in the context of Hybrid MRI/X-ray imaging where transformation of the parallel-beam MRI projections to fan-beam X-ray projections is required. The proposed method is compared to a traditional rebinning method. The results demonstrate that the proposed method is superior to ray-by-ray interpolation and is able to deliver sharper images using the same amount of parallel-beam input projections which is crucial for interventional applications. We believe that this approach forms a basis for further work uniting deep learning, signal processing, physics, and traditional pattern recognition.

Motivation & Objective

• To address the resolution loss inherent in traditional ray-by-ray interpolation during parallel-to-fan beam rebinning in hybrid MRI/X-ray imaging.
• To derive a neural network architecture grounded in analytical signal processing models, integrating known physical operators to reduce parameter count.
• To enable high-frame-rate interventional imaging by learning efficient, compact filters for rebinning without additional hardware.
• To explore whether data-driven filter learning can outperform conventional rebinning while preserving image sharpness.
• To establish a foundation for combining deep learning, physics-based modeling, and signal processing in medical imaging.

Proposed method

• The method formulates the parallel-to-fan beam conversion as a linear system, embedding known operators (e.g., geometric projection models) into the network to reduce unknowns.
• A neural network is derived using the precision learning paradigm, where only the unknown filter components are trained, resulting in a small number of learnable parameters.
• Two filter types are proposed: projection-independent and projection-dependent, both implemented as learnable convolutional filters in the frequency domain.
• The network is trained end-to-end on synthetic MRI projection data, using a mean squared error loss to minimize reconstruction error.
• Regularization via Gaussian smoothing is applied to stabilize filter learning, particularly for projection-independent filters with high amplitude variation.
• The approach replaces the computationally expensive inverse of a system matrix with a learned, efficient convolutional filter in the frequency domain.

Experimental results

Research questions

• RQ1 Can a data-driven neural network architecture be derived from the analytical formulation of the parallel-to-fan beam conversion problem?
• RQ2 Can learning unknown operators in a physics-informed network reduce the number of required training samples and improve reconstruction sharpness?
• RQ3 Does a learned, convolution-based filter outperform traditional ray-by-ray interpolation in terms of image resolution and sharpness?
• RQ4 How do the shape and behavior of projection-dependent and independent filters vary with the number of input projections?
• RQ5 Can regularization strategies such as Lipschitz continuity or symmetry constraints improve filter stability and generalization?

Key findings

• The proposed method produces sharper images than traditional rebinning by eliminating interpolation-induced smoothing.
• The network achieves superior image quality using the same number of parallel-beam projections, which is critical for time-sensitive interventional applications.
• Projection-independent filters show high-amplitude variations in low-projection cases (e.g., 5 or 7 projections), suggesting sensitivity to frequency coverage in training data.
• The projection-dependent filter in the 3- and 5-projection cases exhibits a U-shaped profile, indicating suppression of low-frequency components, which may enable faster MR acquisition.
• Gaussian smoothing regularization led to more stable training and better results than ℓ1 or ℓ2 norms on filter weights or derivatives.
• The results suggest that further reducing the number of required k-space sampling points is possible through improved filter design, potentially enabling analytical determination of optimal sampling trajectories.