Mapping Diffuse Radio Sources Using TUNA: A Transformer-Based Deep Learning Approach

TL;DR

This paper introduces TUNA, a Transformer-based deep learning model that automates and accelerates the detection of faint, diffuse radio sources in astronomical data, outperforming traditional methods in sensitivity and resolution.

Contribution

The paper presents a novel hybrid Transformer-U-Net architecture for radio source segmentation, trained on simulations and applied to real LOFAR data, enabling detection of low surface brightness sources without manual intervention.

Findings

Detects faint diffuse radio sources with high accuracy.

Operates effectively at resolutions 4-6 times lower than input images.

Accelerates discovery in large radio astronomy datasets.

Abstract

Vision Transformers are used via a customized TransUNet architecture, which is a hybrid model combining Transformers into a U-Net backbone, to achieve precise, automated, and fast segmentation of radio astronomy data affected by calibration and imaging artifacts, addressing the identification of faint, diffuse radio sources. Trained on mock radio observations from numerical simulations, the network is applied to the LOFAR Two-meter Sky Survey data. It is then evaluated on key use cases, specifically megahalos and bridges between galaxy clusters, to assess its performance in targeting sources at different resolutions and at the sensitivity limits of the telescope. The network is capable of detecting low surface brightness radio emission without manual source subtraction or re-imaging. The results demonstrate its groundbreaking capability to identify sources that typically require reprocessing at resolutions 4-6 times lower than that of the input image, accurately capturing their morphology and ensuring detection completeness. This approach represents a significant advancement in accelerating discovery within the large datasets generated by next-generation radio telescopes.