Fast reconstruction-based ROI triggering via anomaly detection in the CYGNO optical TPC

This paper presents an unsupervised, reconstruction-based anomaly detection method using a pedestal-trained convolutional autoencoder to efficiently extract Regions of Interest from CYGNO optical TPC images, achieving high signal retention and significant data reduction with low inference latency on consumer hardware.

The Gist

The Big Picture: Finding a Needle in a Haystack (That's Also Moving)

Imagine you are trying to find a specific, rare type of needle in a haystack. But this isn't just any haystack; it's a giant, living haystack that is constantly being photographed by a super-high-definition camera.

Every time the camera takes a picture, it captures the entire haystack. Most of the picture is just empty straw (background noise). But every now and then, a needle (a rare particle event) appears. The problem? The camera is so fast and the pictures are so huge that if you tried to save every single photo, your hard drive would fill up in minutes, and your computer would crash trying to look at them all.

The CYGNO experiment is doing exactly this. They are looking for dark matter particles using a special gas chamber (an Optical Time Projection Chamber). When a particle hits the gas, it creates a tiny, glowing track. But the camera also sees a lot of "static" or "snow" (electronic noise) that looks like a faint, grainy texture.

The Problem: Too Much Data, Too Little Time

The camera takes pictures so fast (about 18 times a second) that the data stream is massive. However, the actual "interesting" part of the picture (the particle track) is tiny—maybe the size of a few grains of sand on a beach.

If you tried to save the whole beach every time a grain of sand moved, you'd waste 99% of your storage space. You need a way to instantly say, "Hey, look here! There's a grain of sand moving!" and throw away the rest of the beach.

The Old Way vs. The New Way

The Old Way (Traditional Physics):
Imagine a team of expert detectives trying to look at every single photo, pixel by pixel, to find the needles. They are very accurate, but they are slow. By the time they finish looking at one photo, the camera has already taken 50 more. They can't keep up with the speed of the camera.

The New Way (Machine Learning):
The authors of this paper built a smart robot (an Artificial Intelligence) that acts like a "noise filter."

1. Training the Robot: First, they showed the robot thousands of photos of the haystack when no needles were present. These are called "pedestal frames." They are just pictures of the static noise. The robot studied these pictures until it memorized exactly what "boring noise" looks like.
2. The Test: Then, they showed the robot new photos that might have needles in them.
3. The Magic: The robot tries to recreate the photo from memory.
   • If the photo is just noise, the robot can recreate it perfectly because it knows that pattern.
   • If there is a needle (a particle track), the robot gets confused. It tries to recreate the noise, but the needle doesn't fit its memory. The robot fails to recreate that specific spot.
   • The Result: The robot highlights the spot where it failed. That failure is the "anomaly." It's the robot saying, "I don't know what this is, so it must be important!"

The Secret Sauce: Teaching the Robot to Say "No"

The paper discovered something very interesting. If you just train a robot to "recreate" the noise perfectly, it sometimes gets too smart. It starts trying to recreate the needles too, thinking they are just part of the noise. This makes it hard to spot the difference.

To fix this, the authors used a clever trick during training:

• They took the boring noise photos and artificially drew fake squiggles and blobs on them (like drawing a smiley face on a blank page).
• They told the robot: "Your job is to recreate the original blank page, not the one with the smiley face."
• The robot had to learn to ignore the fake drawings and focus only on the background noise.

This made the robot much better at spotting the real needles later on, because it was trained to actively reject anything that looked like a structured shape.

The Results: Fast, Tiny, and Accurate

When they tested this system on real data from the CYGNO experiment, the results were impressive:

• Speed: The robot could look at a massive photo and decide what to keep in about 25 milliseconds. That's faster than a human can blink. It's fast enough to keep up with the camera.
• Efficiency: It managed to throw away 97.8% of the picture (the empty background) while keeping 93% of the important signal (the particle tracks).
• Analogy: Imagine you have a 100-page book, but only 3 pages have the story you care about. This system instantly tears out the 97 blank pages and hands you the 3 pages with the story, without losing any of the text.

Why This Matters

This isn't just about saving hard drive space. It's about making future experiments possible. As these detectors get bigger and faster, they will produce so much data that humans can't possibly process it all.

This machine learning approach acts as a smart gatekeeper. It sits right at the entrance of the data stream, filters out the boring stuff in real-time, and only lets the "interesting" parts through for the scientists to study later. It's a lightweight, transparent, and incredibly efficient way to handle the massive data flood of modern physics.

In short: They taught a computer to recognize "boring static" so well that it can instantly spot the "exciting signal" in a sea of noise, saving time, money, and storage space.

Technical Summary

1. Problem Statement

Context: The CYGNO experiment utilizes Optical-readout Time Projection Chambers (TPCs) to search for rare events (e.g., dark matter interactions) in the 1–100 keV energy range. These detectors use scientific CMOS (sCMOS) cameras to capture megapixel-scale images of ionization tracks in a gas mixture (He–CF$_4$).
The Challenge:

• Data Volume: The detectors produce massive data streams (e.g., ~340 MB/s for the CYGNO-04 demonstrator) because the cameras capture the entire active volume.
• Sparsity: The actual physical signals (particle tracks) occupy only a tiny fraction (a few mm$^2$) of the megapixel frames.
• Latency Constraints: Traditional offline reconstruction pipelines are too slow (seconds per frame) for real-time triggering. A solution is needed to filter data at the acquisition rate (~18 frames/s) with a latency budget of ~50 ms.
• Current Gap: Existing unsupervised anomaly detection (AD) methods often focus on global anomaly scores rather than spatial localization, making them less effective for extracting compact Regions of Interest (ROIs) required for data reduction.

2. Methodology

The authors propose an unsupervised, reconstruction-based anomaly detection strategy that operates directly on minimally processed camera frames.

A. Data Strategy: Pedestal-Only Training

• Training Data: The model is trained exclusively on pedestal frames (images taken with the detector's amplification voltage off). These frames contain only intrinsic detector noise (readout noise, fixed patterns, dark counts) and no particle signals.
• Unsupervised Nature: This approach requires no simulation, no labeled signal data, and no fine-grained calibration. The model learns the "normal" noise manifold.

B. Model Architecture

• Network: A convolutional autoencoder with a U-Net-like topology (encoder-decoder with skip connections).
• Input/Output: Processes 1024 $\times$ 1024 pixel images (downscaled from 1525 $\times$ 1525).
• Latent Space: 128-dimensional vector.
• Goal: To reconstruct the noise pattern perfectly while failing to reconstruct localized, structured deviations (particle tracks).

C. Training Objectives (Key Innovation)

The paper compares two training configurations to determine the optimal objective for anomaly localization:

1. Baseline Configuration: Uses a hybrid loss function combining Mean Squared Error (MSE) and Structural Similarity Index (SSIM).
   • Issue: Standard autoencoders tend to "hallucinate" or partially reconstruct faint structured deviations, reducing the contrast in the residual map.
2. Refined Configuration (Proposed): Introduces synthetic perturbations during training.
   • Mechanism: Random "blob" and "track-like" structures are injected into the input images during training.
   • Target: The network is trained to reconstruct the original clean pedestal image, not the perturbed input.
   • Loss Modification: The MSE component of the loss is spatially weighted (up-weighted) in the regions where synthetic perturbations were injected. This forces the network to learn to suppress structured deviations rather than reproduce them.
   • Result: This explicitly regularizes the model to treat structured features as anomalies, sharpening the residual map.

D. ROI Extraction Pipeline

1. Reconstruction: The autoencoder generates a reconstructed image $\hat{x}$.
2. Residual Calculation: Pixelwise residual $r(x) = |x - \hat{x}|$ is computed.
3. Thresholding: A global threshold ($\tau = 0.04$) isolates anomalous pixels.
4. Spatial Aggregation: A morphological closing operation (radius 40 pixels) links nearby anomalous fragments to form coherent ROIs.

3. Key Contributions

1. Pedestal-Only Training: Demonstrated that high-fidelity anomaly detection for optical TPCs can be achieved using only noise-only data, eliminating the need for signal simulation or labeled datasets.
2. Training Objective Design: Identified that the training objective is more critical than architectural complexity. The "refined" strategy (suppressing reconstruction of injected structures) significantly outperforms standard reconstruction objectives and simple pixelwise Gaussian baselines.
3. Transparent Baseline: Established a computationally lightweight, interpretable baseline for online data reduction that is detector-agnostic.
4. Real-Time Feasibility: Validated the approach on real CYGNO prototype data with inference times compatible with real-time triggering.

4. Results

The "Refined Training" configuration was evaluated on 1,563 reconstructed events from the CYGNO prototype.

• Signal Retention: The method retains (93.0 $\pm$ 0.2)% of the reconstructed signal intensity (energy-weighted).
• Data Reduction: It discards (97.8 $\pm$ 0.1)% of the image area, retaining only ~2.2% of the pixels. This represents a reduction of nearly two orders of magnitude in data volume.
• Inference Latency: The average inference time is ~25 ms per frame on a consumer-grade GPU (Apple M1 Pro). This is well within the ~50 ms latency budget required for the CYGNO data acquisition system.
• Comparison:
  • The refined autoencoder outperformed a simple pixelwise Gaussian baseline.
  • A standard autoencoder (without the refined objective) performed similarly to the Gaussian baseline, confirming that naive reconstruction often fails to separate faint signals from noise effectively.
• Energy Dependence: High signal coverage was maintained across the full energy range (O(keV) to hundreds of keV). Cases with near-zero coverage were identified as artifacts of the offline reconstruction (noise fluctuations misidentified as tracks) rather than failures of the ML model.

5. Significance and Impact

• Scalability: As next-generation detectors (like CYGNO-04) scale up to generate hundreds of MB/s, storing full-frame data is unsustainable. This method provides a practical solution for online data reduction, filtering out empty background before storage.
• Physics Preservation: By retaining >90% of the signal intensity while discarding >97% of the data, the method ensures that downstream physics reconstruction algorithms receive high-quality, compressed data without losing critical event information.
• Generalizability: The approach is generic and relies only on the availability of pedestal runs, making it applicable to any optical-readout gaseous TPC experiment.
• Future Outlook: This work lays the foundation for ML-assisted online selection frameworks in future rare-event searches, moving beyond simple triggers to intelligent, reconstruction-aware data filtering.