🔭 astrophysics

YOLO-CIANNA: Galaxy detection with deep learning in radio data: II. Winning the SKA SDC2 using a generalized 3D-YOLO network

This paper presents YOLO-CIANNA, an optimized 3D-YOLO deep learning framework that achieved a record-breaking 9.5% score improvement in the SKA SDC2 competition by delivering a highly pure, accurate, and computationally efficient catalog of 45% more galaxy sources than classical tools from 1TB of simulated radio data.

Original authors: D. Cornu, B. Semelin, P. Salomé, X. Lu, S. Aicardi, J. Freundlich, F. Mertens, A. Marchal, G. Sainton, F. Combes, C. Tasse

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: D. Cornu, B. Semelin, P. Salomé, X. Lu, S. Aicardi, J. Freundlich, F. Mertens, A. Marchal, G. Sainton, F. Combes, C. Tasse

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe is a giant, three-dimensional ocean of radio waves, and hidden within it are thousands of galaxies, each singing a specific "song" at a particular frequency. For decades, astronomers have tried to find these songs, but the ocean is vast, noisy, and filled with static.

Now, imagine the Square Kilometre Array (SKA), a massive new telescope network, is about to turn on. It will generate data so huge and complex that traditional computers would drown trying to process it. To prepare, the SKA organization created a "training camp" called SDC2 (Science Data Challenge 2). They built a perfect, simulated version of the future radio ocean, filled with hidden galaxies, and asked the world's best data scientists to find them.

This paper is the story of how a team called MINERVA (led by researchers in Paris) not only won the training camp but also built a super-smart, self-improving robot to do the job.

Here is the breakdown of their victory, explained simply:

1. The Problem: Finding a Needle in a Haystack (That's on Fire)

The data they were looking at wasn't just a flat picture; it was a 3D cube.

Two dimensions are like a map (Left/Right, Up/Down).
The third dimension is time/frequency (how deep the signal goes).

The galaxies in this cube look like stretched-out, fuzzy clouds. The problem is that the "haystack" is filled with static noise, radio interference from Earth, and leftover echoes from other signals. Finding a galaxy here is like trying to spot a specific whisper in a hurricane while wearing noise-canceling headphones that are slightly broken.

2. The Solution: YOLO-CIANNA (The "Eagle Eye" Robot)

The team used a method called YOLO-CIANNA.

YOLO stands for "You Only Look Once." In computer vision, this is like a security guard who glances at a crowd and instantly spots a person in a red hat, rather than checking every single person one by one.
CIANNA is the custom "brain" (software framework) they built to handle this specific astronomical job.

The Upgrade: In their previous work, the robot could only look at flat 2D pictures. For this challenge, they upgraded the robot's brain to be 3D. Instead of just looking for shapes on a surface, it learned to understand volume. It learned that a galaxy isn't just a dot; it's a 3D structure that stretches across different frequencies, looking a bit like a twisted cigar or a spinning disk.

3. The Secret Sauce: The "Bootstrap" Strategy

Here is where the team got really clever. They faced a tricky problem: How do you teach a robot to find things if you aren't 100% sure what the "real" things look like?

In the training data, there were thousands of faint galaxies. Some were so dim that even the experts weren't sure if they were real or just noise.

The Mistake: If you tell the robot, "Ignore anything faint," it will never learn to find the faint ones. If you tell it, "Find everything," it will get confused by the noise and start hallucinating fake galaxies.

The Fix (The Bootstrap):

Round 1: They trained a "Baby Robot" on the obvious, bright galaxies.
The Test: They let the Baby Robot scan the data again. It found some faint galaxies that the experts had ignored because they were "too risky."
The Lesson: The team said, "Hey, the robot thinks these faint things are real. Let's trust the robot a little bit and add them to the training list."
Round 2: They trained a "Teenager Robot" on the original bright ones plus the new faint ones the Baby found.
Round 3: They repeated this process. Each time, the robot got better at spotting the faint whispers, and the team added those new discoveries to the training list.

This is called bootstrapping. It's like a student who learns the basics, then teaches themselves the advanced stuff by solving problems they thought were too hard, and then using those solutions to learn even more.

4. The Results: A Masterpiece of Efficiency

The results were stunning:

Speed: The robot processed a data cube the size of 1 Terabyte (roughly 200,000 high-definition movies) in just 30 minutes on a single graphics card. A human team would take years.
Accuracy: It found 45% more galaxies than the best traditional tools (like SoFiA) used by other teams.
Purity: It was incredibly accurate. 92% of the galaxies it found were real. When they tuned it to be ultra-careful, it reached 99% purity (only 1 in 100 was a fake).
The Score: They beat the second-place team by a huge margin (13% better). Even after the challenge ended, no one else has managed to beat their score.

5. Why This Matters

Think of this as building the ultimate metal detector for the universe.

Before: Astronomers had to manually sift through data, often missing the faint, distant galaxies that hold the secrets of the early universe.
Now: This AI can scan the entire sky, find thousands of galaxies instantly, and tell us exactly how big they are, how fast they are spinning, and where they are located.

The team is now taking this "metal detector" and testing it on real data from current telescopes (like MeerKAT and ASKAP) before the giant SKA even turns on. If it works there, we are going to discover a universe we never knew existed, all thanks to a robot that learned to listen to the whispers of the cosmos.

In a nutshell: They built a 3D AI that learned to spot galaxies in a noisy radio ocean, taught itself to find the faintest ones by trusting its own discoveries, and did it faster and better than any human or traditional computer ever could.

1. Problem Statement

The Square Kilometre Array (SKA) and its precursors generate massive, high-dimensional, hyperspectral datasets (3D data cubes: 2 spatial dimensions + 1 frequency dimension) that overwhelm classical data processing pipelines.

Specific Challenge: The SKA Observatory Science Data Challenge 2 (SDC2) required the detection and characterization of neutral hydrogen (HI) galaxies within a massive simulated 3D data cube (~1 TB, 5851×5851×6668 voxels).
Key Difficulties:
- High Noise: The data contains thermal noise, radio frequency interference (RFI), and continuum subtraction residues.
- Source Morphology: HI sources are compact in the sky plane but extended across tens of frequency channels, with signal intensity varying non-uniformly.
- Label Scarcity: Supervised learning requires labeled data, but the "truth" catalog for the main challenge data (MAIN cube) was hidden; only a smaller training cube (LDEV) was provided.
- Detection Limits: A vast majority (~85%) of sources in the catalog are too faint to be detected, creating a risk of mislabeling if all catalog entries are used as training targets.

2. Methodology

The authors present YOLO-CIANNA, a regression-based deep learning method adapted from 2D object detection to 3D hyperspectral data.

A. Network Architecture (3D CNN Backbone)

Generalization: The method extends the 2D YOLO-CIANNA framework to 3D by adding a third dimension to the detection grid and output vectors.
Input/Output: The network processes fixed subcubes of 64×64×256 voxels. It outputs a 3D grid (8×8×16) of detection units.
Anisotropic Design: Unlike standard 3D CNNs, the architecture uses anisotropic filters and strides to account for the different signal distributions in spatial vs. spectral dimensions.
- The first layer convolves the frequency axis with large filters (height channels) to capture large spectral structures without prior smoothing.
- The network is deeper (23 layers) but contains fewer parameters (3.35 million) than the previous 2D version, optimized for the specific morphology of HI sources.
Bounding Box Representation: Sources are represented as 3D bounding boxes defined by center coordinates $(x, y, z)$ , dimensions $(w, h, d)$ , and physical parameters (flux, size, line width, position angle, inclination).
Loss Function: Uses a modified Intersection over Union (IoU) metric, specifically DIoU (Distance-IoU), generalized to 3D, to measure box proximity and similarity.

B. Data Preprocessing & Training Strategy

Preprocessing:
- Continuum residuals are masked.
- Channels are normalized by standard deviation to homogenize noise.
- A scaled hyperbolic tangent ( $\tanh$ ) transformation is applied to voxels to enhance contrast and handle dynamic range.
Detectability Selection Function: Since ~85% of cataloged sources are undetectable, the authors defined a selection function based on integrated line flux ( $f$ ) and source volume brightness ( $V_b$ ). This filters the training catalog to include only sources likely to be detectable, preventing the model from learning to detect noise.
Bootstrap Training Strategy:
- Problem: The initial selection function might exclude faint but detectable sources or include undetectable ones, leading to suboptimal training.
- Solution: A multi-step bootstrap process was employed:
  1. Train a BASE model on the initial selected catalog.
  2. Use the BASE model to detect sources in the training data.
  3. Refine the selection function to include fainter sources that the BASE model successfully detected (based on a DIoU match).
  4. Flag difficult sources and retrain (BT1, BT2, BT3).
- This iterative process allowed the model to progressively learn to detect fainter sources without being overwhelmed by noise.

C. Inference Pipeline

The full MAIN cube is decomposed into overlapping subcubes (overlap of 8 pixels in sky, 32 channels in frequency) to ensure sources are fully contained.
Non-Maximum Suppression (NMS): A two-stage NMS is applied:
1. Within each input subcube (3D DIoU threshold = 0.1).
2. Across overlapping inputs (3D DIoU threshold = -0.3).
Threshold Optimization: An automated search optimizes the objectness confidence threshold per frequency bin to maximize the final score.

3. Key Contributions

First 3D Generalization of YOLO-CIANNA: Successfully adapted a 2D regression-based detection framework to 3D hyperspectral radio data, handling the unique anisotropy of HI emission.
Custom 3D CNN Backbone: Designed a specialized architecture that outperforms generic pre-trained models and standard 2D-backbone approaches for this specific task.
Bootstrap Training Framework: Introduced a novel iterative strategy to refine the training sample selection function, enabling the detection of sources fainter than the initial theoretical limits.
State-of-the-Art Performance: Achieved the highest score in the SDC2 challenge, significantly outperforming both classical tools (SoFiA) and other deep learning approaches.

4. Results

Challenge Score: The final BT1 model achieved a score of 25,453, improving the original challenge submission (MINERVA v0.1) by 9.5% and placing 13.2% higher than the second-best team.
Detection Performance:
- Purity: 92.3% (99.1% when optimized for high purity).
- Completeness: The method detected 45% more confirmed sources than concurrent classical tools.
- Characterization: Achieved best-in-class accuracy for source properties (flux, size, line width, etc.).
Efficiency: The method processed the entire ~1 TB data cube in 30 minutes on a single GPU (RTX 6000 Ada), demonstrating high computational efficiency.
Robustness: The model successfully detected sources fainter than the training selection limit, proving its ability to generalize beyond the strict training distribution.

5. Significance

Validation of 3D CNNs: The results confirm that 3D Convolutional Neural Networks are superior to classical algorithms and 2D approaches for extracting complex 3D patterns in noisy hyperspectral data.
Readiness for SKA: The method represents a critical step toward handling the data deluge from the SKA and its precursors (e.g., LADUMA, WALLABY). The authors are already applying this to real observational data.
Open Science: The authors have made all code, trained models, and detection catalogs publicly available, facilitating future comparisons and the development of next-generation tools.
Methodological Insight: The paper highlights that successful application of deep learning in astronomy requires tailored architectures and careful handling of training data biases (via bootstrapping), rather than simply applying off-the-shelf computer vision models.

In conclusion, YOLO-CIANNA v1.0 sets a new benchmark for automated galaxy detection in radio astronomy, offering a fast, accurate, and scalable solution for the upcoming era of large-scale radio surveys.