Radio U-Net: a convolutional neural network to detect diffuse radio sources in galaxy clusters and beyond

TL;DR

This work addresses the challenge of detecting faint, diffuse radio sources in large interferometric surveys by introducing Radio U-Net, a modified U-Net that yields per-pixel probabilities for diffuse emission. Trained on synthetic LOFAR-like observations derived from cosmological simulations, the network can segment complex morphologies and support fast, automated detection across large datasets. When applied to real LoTSS-DR2/PSZ2 cluster data, it achieves about $73 ext{\%}$ accuracy with $83 ext{\%}$ recall at an optimal threshold, while preserving morphology in low-quality images; fine-tuning on actual data yields limited improvements due to dataset size. The method demonstrates the viability of transferring knowledge from high-fidelity simulations to real observations and highlights the potential for HPC-enabled, automated surveys in the era of the SKA, with future work focusing on more realistic training sets and multi-wavelength integration.

Abstract

The forthcoming generation of radio telescope arrays promises significant advancements in sensitivity and resolution, enabling the identification and characterization of many new faint and diffuse radio sources. Conventional manual cataloging methodologies are anticipated to be insufficient to exploit the capabilities of new radio surveys. Radio interferometric images of diffuse sources present a challenge for image segmentation tasks due to noise, artifacts, and embedded radio sources. In response to these challenges, we introduce Radio U-Net, a fully convolutional neural network based on the U-Net architecture. Radio U-Net is designed to detect faint and extended sources in radio surveys, such as radio halos, relics, and cosmic web filaments. Radio U-Net was trained on synthetic radio observations built upon cosmological simulations and then tested on a sample of galaxy clusters, where the detection of cluster diffuse radio sources relied on customized data reduction and visual inspection of LOFAR Two Metre Sky Survey (LoTSS) data. The 83% of clusters exhibiting diffuse radio emission were accurately identified, and the segmentation successfully recovered the morphology of the sources even in low-quality images. In a test sample comprising 246 galaxy clusters, we achieved a 73% accuracy rate in distinguishing between clusters with and without diffuse radio emission. Our results establish the applicability of Radio U-Net to extensive radio survey datasets, probing its efficiency on cutting-edge high-performance computing systems. This approach represents an advancement in optimizing the exploitation of forthcoming large radio surveys for scientific exploration.

Figures (9)

• Figure 1: Schematic representation of the Radio U-Net architecture.
• Figure 2: Left panels: two examples of sky images with a contour level at $10^{-8}$ Jy/pixel, which represents the reference mask used to train the network. Central panels: clean images with contours drawn at 3$\sigma$, with $\sigma=2\times10^{-6}$ Jy/beam for the top image and $\sigma=1.5\times10^{-5}$ Jy/beam for the one at the bottom. The restoring beam is $6\arcsec$. Right panels: probability images created by Radio U-Net with contours at 0.5.
• Figure 3: Scheme of the tiling procedure. During training, the tiles are all independently processed. The categorical cross-entropy loss function, which is the metric used during the training, is computed for each tile separately. When the evaluation program is used to apply the trained network to new images, the tiles are overlapped. They are still processed independently and the central cut-outs are reassembled to avoid boundary problems. The size of the figures can vary but the tile size is fixed to 192 pixels.
• Figure 4: Examples of LoTSS-DR2/PSZ2 galaxy clusters processed by Radio U-Net and successfully classified on the basis of their $\mathcal{R}$-value (see Sec. \ref{['sec:results']}). Low-resolution ($20\arcsec$ beam) images of the LoTSS survey are shown in the first row while their probability maps are shown in the bottom row. The 0.5 probability contour is also reported in the images of the first row. The white circle shows the reference 2.2$R_{500}$ radius where $\mathcal{R}$ is computed. NDE marks a non-detection, RH and RR mark the presence of a radio halo and a radio relic, respectively. U is used to classify an uncertain detection that, in this case, is a radio-halo.
• Figure 5: Examples of LoTSS-DR2/PSZ2 galaxy clusters processed by Radio U-Net (see Sec. \ref{['sec:results']}). Low-resolution ($20\arcsec$ beam) images of the LoTSS survey are shown in the first row while their segmented maps are shown in the bottom row. The white circle shows the reference 2.2$R_{500}$ radius where $\mathcal{R}$ is computed. The 0.5 probability contour is also reported in the images of the first row. NDE marks a non-detection, RH marks the presence of a radio halo while U is used to classify an uncertain detection. The first column shows a cluster incorrectly classified as detected. The second column shows a case where the cluster is correctly classified as detected but the probability map is dominated by the central extended radio galaxy rather than by the radio halo. The third column shows a case of a radio halo not detected by Radio U-Net.