Tuna-R: Deep Learning for Diffuse Radio Segmentation

• Tuna-R is a deep-learning framework for detecting and segmenting faint, diffuse radio emission in interferometric surveys with exceptional scalability and sensitivity.
• It leverages a hybrid TransUNet backbone that combines convolutional layers and transformer modules to capture both local spatial details and global contextual features from LOFAR data.
• Performance metrics show Tuna-R outperforms traditional U-Nets—with recall up to 0.62 and improved IoU—thereby streamlining automated radio source detection.

Tuna-R is a deep-learning framework for the detection and segmentation of extremely faint, diffuse radio emission in interferometric sky surveys, built upon a vision transformer–augmented U-Net architecture. Developed for high-throughput radio astronomy applications, Tuna-R enables automated identification of low surface-brightness structures, such as galaxy cluster megahalos, bridges, and filaments, from native-resolution interferometric images without need for manual source subtraction or re-imaging. Its unprecedented scalability and sensitivity are tailored for next-generation radio sky surveys, offering robust generalization to real LOFAR data with significant calibration and imaging artifacts (Sanvitale et al., 15 Jul 2025).

1. Hybrid Architecture: TransUNet for Radio Astronomy

Tuna-R is underpinned by the TransUNet backbone, a hybrid encoder–decoder architecture combining convolutional layers with transformer modules to integrate both local spatial and global contextual information. The encoder uses a ResNet-50 convolutional network pretrained on ImageNet21k as its backbone, which processes the $H \times W \times C$ input image and produces a feature map of dimensions $H'=H/16$, $W'=W/16$, and channel depth $C'$ (e.g., $C'=2048$). This downsampled feature map is partitioned into $N = H' \times W'$ patches (with patch size $P_s = 1$), each embedded into a $D$-dimensional latent space ($D=768$). Positional encodings are added to the input sequence, which is then processed through $L=12$ transformer layers, each with $H'=H/16$0 attention heads of dimension $H'=H/16$1.

The transformer output sequence is reshaped back to spatial dimensions and fed into a U-Net decoder, which employs cascaded upsampling (2$H'=H/16$2 per block) and $H'=H/16$3 convolutions, with skip connections from encoder layers preserving fine-scale spatial detail. This modular architecture efficiently captures global and local features critical for segmenting diffuse extragalactic sources (Sanvitale et al., 15 Jul 2025).

2. Training Datasets and Pipeline

Tuna-R is initially trained on synthetic datasets derived from cosmological magnetohydrodynamic (MHD) simulations produced using Enzo, covering a $H'=H/16$4 volume with 41.65~kpc cell size and primordial $H'=H/16$5~nG. Synchrotron emission at $H'=H/16$6 is computed via shock acceleration prescriptions and projected into light cones spanning $H'=H/16$7, yielding $H'=H/16$8 mock skies. For each realization, LOFAR-HBA imaging systematics—including beam convolution, primary-beam correction, and Gaussian noise (rms $H'=H/16$9~mJy/beam)—are injected using WSClean and custom routines.

Ground-truth binary masks assign pixels above $W'=W/16$0~Jy/pixel to the diffuse class ($W'=W/16$1\% of the area). Two key dataset variants are constructed at LOFAR's native $W'=W/16$2 ($W'=W/16$3 beam) and $W'=W/16$4 (convolved) resolutions. Application fields in the LOFAR Two-meter Sky Survey—both raw and post-source-subtraction—allow for robust evaluation on real survey data (Sanvitale et al., 15 Jul 2025).

Training optimizes a combined cross-entropy and Dice loss ($W'=W/16$5) over $W'=W/16$6 epochs using the Adam optimizer (learning rate $W'=W/16$7, batch size $W'=W/16$8, tile size $W'=W/16$9), with no performance gain found for minority-class weighting. Five independent runs are used to estimate uncertainty and robustness.

3. Quantitative Performance and Evaluation

Performance is systematically evaluated using recall, precision, intersection-over-union (IoU), and pixel-wise accuracy on both simulated and real datasets. The key metrics, averaged over multiple test fields, are summarized below.

Dataset	Recall	Precision	IoU	Accuracy
6'' TUNA	0.50±0.01	0.61±0.02	0.38±0.01	0.97±0.01
20'' TUNA	0.62±0.02	0.76±0.02	0.52±0.01	0.97±0.01
6'' R-UNet	0.39±0.01	0.64±0.01	0.30±0.01	0.96±0.01
20'' R-UNet	0.53±0.01	0.73±0.01	0.43±0.01	0.97±0.01

On real LOFAR survey data, Tuna-R achieves mean recall and IoU (for extended, diffuse emission) of $C'$0 and $C'$1 respectively, significantly outperforming state-of-the-art U-Nets (recall $C'$2, IoU $C'$3) (Sanvitale et al., 15 Jul 2025). The network robustly detects emission down to brightness thresholds near $C'$4~Jy/beam in LoTSS-DR2, exceeding $C'$5\% completeness at these sensitivity floors.

4. Application to Science Cases and Inference Pipeline

Tuna-R is deployed for end-to-end segmentation on both mock and real LOFAR data. The complete inference pipeline consists of:

• Reading calibrated FITS images
• Log-scaling and normalizing flux to $C'$6
• Tiling into $C'$7 sub-images with $C'$8–$C'$9\% overlap
• GPU-based model evaluation to produce confidence maps
• Inverse normalization and mosaicking back to the full field for output mask generation as FITS

Empirical runtime benchmarks on an NVIDIA A100 (Leonardo supercomputer) demonstrate processing of a $C'=2048$0~pixel field at $C'=2048$1 resolution in $C'=2048$2~seconds. This contrasts with human-involved, day-scale re-imaging procedures using classical pipelines.

The network reliably identifies science targets including the 3~Mpc A399–A401 radio bridge and multiple $C'=2048$3~Mpc megahalos, as validated by recovery of morphology at reference resolutions $C'=2048$4–$C'=2048$5 lower than the input. Tuna-R eliminates the need for explicit discrete-source subtraction and retapered imaging, generalizing to complex calibration artefacts and noise in survey data (Sanvitale et al., 15 Jul 2025).

5. Comparison to Classical Segmentation and Impact

Traditional source identification for diffuse radio emission has relied on manual or semi-automated pipelines involving compact source subtraction, re-imaging at reduced angular resolution (e.g., $C'=2048$6–$C'=2048$7 for LOFAR), and thresholding, incurring substantial computational (days per field) and human time investment. Tuna-R replaces these with a parallelized, fully automated workflow, achieving recall $C'=2048$8 and IoU $C'=2048$9 over state-of-the-art U-Nets, and completeness matching or exceeding the low-resolution, reprocessed legacy maps.

Unlike pure CNN-based architectures, Tuna-R’s hybrid attention mechanism preserves both compact and extended structures, recovers faint emission at the sensitivity limit, and is robust against artefacts and residual sidelobe confusion. End-to-end inference not only accelerates the science pipeline but also enables blind, scalable surveys for rare objects such as megahalos, relics, and filamentary bridges (Sanvitale et al., 15 Jul 2025).

6. Limitations and Future Prospects

Current limitations arise from patch-based attention effects, introducing mild smoothing and $N = H' \times W'$0\% brightness-dependent biases at intermediate intensities; at $N = H' \times W'$1–$N = H' \times W'$2~Jy/arcsec$N = H' \times W'$3 the relative bias $N = H' \times W'$4, rising to a $N = H' \times W'$5 overestimate at the highest surface brightnesses (Sanvitale et al., 19 Oct 2025). The architecture is optimized for native-resolution ($N = H' \times W'$6) input; when working at coarser resolution, risk of confusion by unresolved point sources increases. Generalization to other frequency ranges and instruments (e.g. MWA, SKA-LOW) is a target for future work, as is integration with flux regression to provide quantitative brightness estimates and uncertainty-aware object catalogs.

Advancements include augmenting training with realistic AGN/radio-galaxy populations, leveraging uv-plane data, and developing network-internal flux scaling. Tuna-R is positioned for adoption in upcoming SKAO-era petascale workflows and as a foundational engine for next-generation all-sky, faint radio source discovery (Sanvitale et al., 15 Jul 2025).

---

Key references:

• "Mapping Diffuse Radio Sources Using TUNA: A Transformer-Based Deep Learning Approach" (Sanvitale et al., 15 Jul 2025)
• "Estimating Flux Densities of Diffuse Cosmological Radio Sources Exploiting Vision Transformers" (Sanvitale et al., 19 Oct 2025)