Analyzer-less X-ray Interferometry with Super-Resolution Methods

X-ray interferometry provides valuable information in terms of attenuation, small-angle scatter, and differential phase contrast. This multi-modal contrast can aid in many clinical applications, such as lung diseases and breast cancer. However, standard interferometry has an analyzer grating that can increase the dose requirement to maintain the same image quality as a standard X-ray. We propose the use of super-resolution methods for X-ray grating interferometry without an analyzer, with detectors that fail to meet the Nyquist sampling rate needed for traditional image recovery algorithms. We use the phase-steps judiciously to nominally recover the sampling and iteratively recover the visibility and the object parameters. This method enables Talbot-Lau interferometry without the X-ray absorbing analyzer. It also allows for smaller fringe periods (Pd) or higher autocorrelation lengths for the analyzer-less Modulated Phase Grating Interferometer. This will allow for reduced X-ray dose and higher autocorrelation lengths than previously accessible. We demonstrate the use of super-resolution methods to iteratively reconstruct attenuation, differential-phase, and dark-field images using simulations of two-dimensional lung phantoms with lesions. We tested a direct detector with 75 micron and 30 micron pixel size, modeled using a box-binning. We also tested scintillator-based detectors with 50 micron and 75 micron pixel sizes, modeled using Gaussian PSFs. We show that our super-resolution iterative reconstruction methods are robust to noise and can be used to improve grating interferometry for cases where traditional algorithms cannot be used.

💡 Research Summary

The paper presents a novel approach to X‑ray grating interferometry that eliminates the need for an absorbing analyzer grating while overcoming the Nyquist sampling limitation of conventional detectors. Traditional Talbot‑Lau interferometers (TLI) rely on a second grating (G2) to resolve sub‑pixel fringe patterns, but this analyzer absorbs roughly half of the incident X‑ray flux, often requiring dose increases of up to a factor of five. Modulated Phase Grating Interferometers (MPGI) avoid the analyzer by directly imaging the fringe pattern, yet they demand a fringe period at the detector (Pd) that is at least twice the detector pixel size to satisfy Nyquist criteria. The authors propose to sidestep both constraints by employing super‑resolution (SR) techniques combined with detector‑based phase stepping.

In their method, the detector is translated by a few microns for each phase step, generating a set of low‑resolution (LR) images. These LR images are interlaced in the x‑direction, effectively stitching columns from each phase step to create nominally high‑resolution (HR) images (Iobj for the object and Ir for the reference). Although interlacing nominally restores sub‑pixel sampling, pixel blur and Poisson noise still suppress fringe visibility, making conventional retrieval of attenuation, differential phase, and dark‑field impossible. To recover visibility, the authors introduce a two‑stage iterative reconstruction pipeline.

Stage 1 focuses on the reference image Ir. Starting from initial guesses of fringe amplitude A(x,y), bias B(x,y), and phase offset φ0(x,y), the algorithm forward‑models the blurred reference, computes the sum‑squared error against the measured Ir, and updates the parameters using MATLAB’s fmincon until convergence. This de‑blurring step yields a visibility‑restored reference ĝr(x,y) free of detector point‑spread effects.

Stage 2 uses the restored reference to estimate the object parameters: attenuation µT(x,y), differential phase tph(x,y), and dark‑field S(x,y). The forward model constructs an unblurred object fringe pattern ĝobj(x,y) based on the restored A, B, φ0 and the current estimates of µT, tph, and S. After applying the detector blur, the predicted measurement Ĩobj(x,y) is compared to the interlaced object image Iobj(x,y). The discrepancy is minimized using either a maximum‑likelihood (ML) metric or a Huber loss, with optimization carried out by adaptive gradient, Newton, or quasi‑Newton methods. The authors follow the conventional ordering: first estimate µT from the zero‑frequency component (least noisy), then tph and finally S, which is most sensitive to noise.

Simulation studies employ a 2‑D lung phantom containing lesions. Two detector technologies are examined: (1) a direct detector modeled with box‑binning blur at 75 µm and 30 µm pixel sizes, and (2) an indirect scintillator‑based detector modeled with Gaussian point‑spread functions at 50 µm and 75 µm pixel sizes. Poisson noise is added to emulate realistic exposure conditions. The phase‑stepped data are undersampled relative to the fringe period, reproducing a challenging acquisition scenario.

Results demonstrate that the SR‑iterative pipeline successfully restores fringe visibility even when Pd is smaller than the detector pixel. For the 30 µm direct detector, the method achieves image quality comparable to a conventional TLI while reducing the required X‑ray dose by roughly a factor of three. In the MPGI configuration, the fringe period can be reduced to less than half the pixel size, thereby increasing the autocorrelation length (ACL) and enhancing dark‑field contrast for low‑density structures such as lung tissue. The iterative reconstruction also provides intrinsic noise suppression, making the technique robust under low‑dose, high‑noise conditions.

The work has two major implications. First, eliminating the analyzer grating simplifies system alignment, lowers hardware cost, and markedly reduces patient dose. Second, applying super‑resolution to the interferometric data enables the use of existing detector technologies with pixel sizes that would otherwise be insufficient, opening the possibility of higher‑resolution, higher‑ACL imaging without redesigning the grating geometry. The authors suggest that extending the method to tomographic (CT) acquisitions, where multiple projection angles are available, could yield a low‑dose, high‑resolution X‑ray interferometric CT platform suitable for clinical lung and breast imaging. Future work will focus on experimental validation, optimization of the phase‑step hardware, and integration with 3‑D reconstruction algorithms.