Radio interferometers and wide-area continuum surveys are producing increasingly large and complex imaging datasets, making manual source identification and classification impractical. Accurate, fast and fully automated methods are therefore required to detect and characterise two distinct but related populations: extended radio sources associated primarily with active galactic nuclei, including jets and radio lobes, and diffuse low-surface-brightness sources, such as radio halos, relics, bridges and filamentary emission associated with galaxy clusters and the cosmic web. Both classes exhibit complex, irregular and frequently multi-component morphologies that are difficult to recover with conventional source-finding algorithms.
TUNA (TransUNet for Astrophysical data) has been introduced as a deep-learning framework for the automated segmentation of radio sources. Its effectiveness has already been investigated across all the targeted source classes, including extended radio galaxies and diffuse halos, relics and bridges. TUNA is based on a hybrid TransUNet architecture combining a convolutional encoder, a Vision Transformer and a U-Net-like decoder. Convolutional layers extract local brightness and morphological features, while multi-head self-attention captures long-range spatial dependencies between image regions. Skip connections preserve fine spatial information during the reconstruction of full-resolution pixel-level segmentation maps.
The TUNA training and inference pipelines are designed to exploit GPU acceleration, multi-GPU data parallelism and mixed-precision computation. Computationally intensive convolutional and attention operations are executed on GPUs, while the workload can be distributed across multiple devices. Mixed-precision execution effects on memory and power consumption, and computational performance is being investigated.
This contribution will present the application of TUNA to the systematic processing of LOFAR Two-metre Sky Survey Data Release 3 pointings. The talk will describe the processing workflow and assess the method from both scientific and computational perspectives. Its effectiveness in detecting and segmenting extended and diffuse radio sources will be evaluated together with its computational efficiency, scalability, GPU utilisation and multi-GPU performance. Scientific accuracy and computational performance will be presented and discussed as complementary requirements for the deployment of AI-based source-segmentation methods in current and future large-area radio surveys.