POLISH'ing the Sky: Wide-Field and High-Dynamic Range Interferometric Image Reconstruction with Application to Strong Lens Discovery

TL;DR

This work builds upon and extends the POLISH framework, a recent DL model for radio interferometric imaging, and introduces key improvements to enable robust reconstruction and super-resolution under real-world conditions: a patch-wise training and stitching strategy for scaling to wide-field imaging and a nonlinear arcsinh-based intensity transformation to manage high dynamic range.

Abstract

Radio interferometry enables high-resolution imaging of astronomical radio sources by synthesizing a large effective aperture from an array of antennas and solving a deconvolution problem to reconstruct the image. Deep learning has emerged as a promising solution to the imaging problem, reducing computational costs and enabling super-resolution. However, existing DL-based methods often fall short of the requirements for real-world deployment due to limitations in handling high dynamic range, large field of view, and mismatches between training and test conditions. In this work, we build upon and extend the POLISH framework, a recent DL model for radio interferometric imaging. We introduce key improvements to enable robust reconstruction and super-resolution under real-world conditions: (1) a patch-wise training and stitching strategy for scaling to wide-field imaging and (2) a nonlinear arcsinh-based intensity transformation to manage high dynamic range. We conduct comprehensive evaluations using the T-RECS simulation suite with realistic sky models and point spead functions (PSF), and demonstrate that our approach significantly improves reconstruction quality and robustness. We test the model on realistic simulated strong gravitational lenses and show that lens systems with Einstein radii near the PSF scale can be recovered after deconvolution with our POLISH model, potentially yielding 10$\times$ more galaxy-galaxy lensing systems from the Deep Synoptic Array (DSA) survey than with image-plane CLEAN. Our results highlight the potential of DL models as practical, scalable tools for next-generation radio astronomy.

Figures (9)

• Figure 1: Illustration of patch-wise measurements and cross-patch contamination in wide-field interferometric imaging. Left: the full–field-of-view dirty measurement ($12,960\times12,960$ pixels), with the highlighted region indicating the patch under consideration ($1,296\times1,296$ pixels). Top right: the ground-truth sky patch. Top middle: the corresponding patch extracted from the full-FOV dirty measurement. Bottom middle: a simulated measurement obtained by convolving the ground-truth patch alone with the PSF. Bottom right: the corresponding patch of the CLEAN reconstruction from the full-FOV measurement. The measurement patch extracted from the full FOV contains strong artifacts induced by bright sources outside the patch, visible as extended streaks from PSF sidelobes (e.g., from sources below the patch). In contrast, these nonlocal artifacts are absent when simulating the measurement from the ground-truth patch alone, highlighting that the standard forward model does not apply locally at the patch level and motivating learning-based approaches that can handle such cross-patch contamination.
• Figure 2: Visualization of the nonlinear transform used for dynamic range reduction.(a)$\mathsf{AsinhStretch}(x; a)$ defined in (\ref{['eqn:asinh']}) with $a \in \{0.1, 0.01, 0.001\}$. (b) Visualize of the dynamic range distribution of one test image before (top) and after (bottom) applying $\mathsf{AsinhStretch}(x; a=0.001)$ . One can see that the maximum dynamic range is reduced by over one order of magnitude after applying the transformation.
• Figure 3: Visual comparison of galaxy detection results across different reconstruction methods. We intentionally show detections only in the top half of each image to facilitate comparison of both the detections and the underlying image quality. Top row: ground truth sky (48 galaxies), ground truth with added Gaussian noise, and the corresponding low-resolution measurement. Bottom row: reconstructions obtained with CLEAN, POLISH, and POLISH++. Detected sources are marked by ellipses, with true positives shown in green, false positives in red, and false negatives in blue (reported as TP/FP/FN in the caption of each method). For each detection, the major and minor axis lengths of the ellipse are set to twice the corresponding FWHM values. The cyan zoom-in regions highlight the improved spatial resolution achieved by POLISH+ and POLISH++ relative to the blurred CLEAN reconstructions, demonstrating their super-resolution capability. The green zoom-in region illustrates background artifacts introduced by POLISH, whereas POLISH++ yields a markedly cleaner, artifact-free reconstruction with reduced hallucinations. The top colorbar corresponds the pixel intensity scale of the "Ground truth + noise" and the "Measurement" panels, while the bottom colorbar corresponds to that of the other four panels.
• Figure 4: Detection performance versus progressively larger signal-to-noise ratio (SNR) threshold. From left to right, the panels show precision, recall, and $F_1$ score for CLEAN, POLISH, POLISH+, and POLISH++ as the SNR threshold increases from 3 to 300,000. For each subplot, the leftmost point roughly matches the corresponding values in Tab. \ref{['tab:baselines']} as almost all sources have flux $>3$. The background histogram in the rightmost panel (gray, logarithmic scale) visualizes the SNR distribution of all sources, indicating that most sources have relatively low SNRs. Overall, POLISH+ and POLISH++ substantially outperform the baselines. Furthermore, POLISH++ achieves a generally better tradeoff between precision and recall, leading to an improved $F_1$ score in the low-SNR-threshold regime which includes the majority of sources.
• Figure 5: Comparison of galaxy property estimation accuracy between CLEAN (blue) and POLISH++ (orange). The scatter plots compare the predictions (vertical axes) versus the true values (horizontal axes) for major axis FWHM $\theta_A$ (left), minor axis FWHM $\theta_B$ (middle), and flux (right), across all true positive detections with an SNR threshold of 300. The dashed line indicates perfect prediction. POLISH++ produces shape estimates that distribute closer to the diagonal, while CLEAN provides more accurate flux estimates. It is also clear that CLEAN’s resolving power asymptotes to the PSF angular scale, while POLISH achieves accurate super-resolution below the intrinsic resolution of the PSF ($\approx3.3"$).