🔬 physics

Computer vision-based neural networks for radioisotope identification in urban environments

This paper proposes a computer vision-based approach that converts list-mode gamma-ray data into multi-channel spectrograms to leverage neural networks for improved radioisotope identification in urban environments, demonstrating that a convolutional neural network outperforms traditional non-negative matrix factorization methods at low false positive rates on the RADAI benchmark.

Original authors: Masen Bachleda, Peter Lalor

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Masen Bachleda, Peter Lalor

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 you are walking through a busy, noisy city square trying to find a specific person wearing a bright red hat. The problem is that the city square is full of other people wearing red hats (background noise), and the person you are looking for might only flash their hat for a split second before disappearing into the crowd. This is essentially what scientists face when trying to find hidden radioactive materials in a city using gamma-ray detectors.

Here is a breakdown of the paper's work, explained simply:

The Challenge: Finding a Needle in a Haystack

Radioactive materials emit energy that detectors can "see" as a spectrum (a graph of energy levels). In a city, however, the ground, buildings, and concrete naturally emit their own radiation. This creates a constant, shifting "hum" of background noise.

The Problem: Traditional methods look for specific "peaks" in the data, like looking for a specific shape in a cloud. But in a moving city, the background changes constantly, and the radioactive source might only be there for a few seconds. It's like trying to spot a friend in a crowd while you are both running, and the crowd keeps changing its clothes.
The Data: The researchers used a massive, computer-generated dataset called RADAI. It simulates a detector driving through a virtual city for hours, encountering rare radioactive sources amidst a sea of natural background radiation.

The New Idea: Turning Sound into a Movie

Instead of looking at the data as a single static graph, the authors decided to treat it like a movie.

The Waterfall: They took the raw data and stacked it up over time to create a "waterfall spectrogram." Imagine a waterfall where the horizontal axis is the type of energy (like musical notes) and the vertical axis is time.
The Analogy: Think of a standard photo. A color photo has three layers (Red, Green, Blue) to create a full picture. The authors treated their "waterfall" similarly. Instead of stacking time on top of each other as separate rows, they treated consecutive moments in time as color channels.
- Just as your eyes combine Red, Green, and Blue to see a full image, the computer combines these "time slices" to see a pattern that includes both what the energy is and how it changes over time.

The Test: Three Different "Eyes"

The team trained three different types of computer vision "eyes" (neural networks) to look at these waterfall movies and spot the radioactive sources:

The MLP (Multilayer Perceptron): A basic brain that looks at the whole picture at once without understanding how parts relate to each other.
The CNN (Convolutional Neural Network): A sophisticated eye that looks for local patterns, like recognizing a specific shape or texture, similar to how we recognize a face by looking at eyes and a mouth together.
The ViT (Vision Transformer): A high-tech eye that can look at the whole picture and understand how distant parts relate to each other (like connecting a nose on the left side of a face to a chin on the right).

The Results: Who Won the Game?

The researchers tested these "eyes" against the current gold standard, a method called NMF (Non-Negative Matrix Factorization), which is like a very strict librarian who only recognizes items if they perfectly match a pre-written list.

When the rules were slightly relaxed (allowing 1 false alarm per hour):
The CNN won. It was better at spotting the radioactive sources than the librarian (NMF). It found more sources and identified them more accurately.
- Why? The CNN learned to recognize subtle patterns and changes over time that the librarian's rigid list missed. It was like a detective who notices a person's gait or how they move, not just their face.
When the rules were very strict (allowing very few false alarms):
The Librarian (NMF) won again. When the system was forced to be extremely cautious to avoid false alarms, the neural networks started missing things, while the template-based method held its ground.
- Why? The neural networks are great at finding complex, hidden patterns, but they sometimes get "too excited" and see things that aren't there when the bar for accuracy is set incredibly high.

The Breakdown by Source Type

The paper also looked at what kind of radioactive material was being found:

Industrial Sources (like 137Cs): These have very clear, distinct signatures. The "Librarian" (NMF) was actually better at these because they are easy to match to a list.
Nuclear Materials & Natural Background (NORM): These are the hardest to find because they look very similar to the city's natural background. Here, the CNN was a massive improvement, finding these sources 20 times better than the Librarian. It learned to spot the tiny, subtle differences that human-designed lists couldn't catch.

The Bottom Line

This paper shows that by turning radiation data into "movies" and using computer vision techniques (specifically CNNs), we can build better detectors for finding hidden radioactive materials in cities.

The Good: These new AI methods are much better at finding difficult, hard-to-spot sources that look like background noise.
The Catch: They still struggle a bit when the rules demand zero mistakes.
The Future: The authors suggest that the best solution might be a hybrid team: using the "Librarian" for easy-to-spot items and the "AI Detective" for the tricky, hidden ones.

The paper concludes that while AI isn't perfect yet, it is already beating traditional methods in the most challenging scenarios, proving that teaching computers to "watch" radiation data like a movie is a powerful new strategy.

Technical Summary: Computer Vision-Based Neural Networks for Radioisotope Identification in Urban Environments

Problem Statement
The identification of radioisotopes in mobile urban search scenarios presents significant challenges for automated detection algorithms. Urban environments are characterized by non-uniform backgrounds, including naturally occurring radioactive material (NORM) in building materials and soil, which create location-dependent and temporally varying background radiation. This variability, combined with the transient nature of source encounters (brief events within continuous measurements) and severe class imbalance (where background samples vastly outnumber rare threat signatures), obscures threat signatures. Traditional peak-based analysis methods struggle in these dynamic, short-duration scenarios, while existing machine learning approaches often treat classification as a static pattern recognition problem on accumulated spectra, discarding valuable temporal information. Furthermore, the development of robust algorithms is hindered by the scarcity of large, labeled real-world datasets due to safety constraints and the difficulty of acquiring ground truth.

Methodology
To address these challenges, the authors propose a machine learning framework that converts raw list-mode gamma-ray data into two-dimensional waterfall spectrograms and applies computer vision architectures to these images. The core innovation lies in the data representation: rather than treating the waterfall as a conventional spatial image, the authors employ a representation where consecutive time spectra form the input channels, analogous to RGB channels in color images. This design encodes both spectral information (along the energy axis) and temporal information (across the channels), enabling neural networks to learn patterns that distinguish source signatures from background fluctuations.

The methodology utilizes the Radiological Anomaly Detection and Identification (RADAI) benchmark dataset, a large-scale synthetic dataset featuring high-fidelity Monte Carlo simulations of mobile detection scenarios with detailed ground-truth labels. The preprocessing pipeline involves:

Waterfall Generation: Binning photon events into 1D gamma spectra using a nonlinear (square-root) energy binning scheme. These spectra are stacked vertically to form 2D waterfalls, where the vertical axis represents time progression and the horizontal axis represents spectral energy.
Sliding Window: A sliding window approach is used to generate heavily overlapping samples, increasing the effective training set size and capturing source encounters at various temporal positions within the waterfall.
Architectures: Three computer vision architectures are evaluated:
- Multilayer Perceptron (MLP): Flattens the waterfall to test if learned feature relationships alone suffice.
- Convolutional Neural Network (CNN): Uses 1D convolutions along the energy axis, treating time windows as input channels to learn consistent spectral patterns (e.g., photopeak shapes) across time.
- Vision Transformer (ViT): Adapts self-attention mechanisms to capture long-range correlations between distant spectral features.
Training and Optimization: Models are trained using Focal Loss to address class imbalance. The primary optimization metric is the True Identification Rate (TIDR), constrained by False Positive Rate (FPR) targets (e.g., < 1 false alarm per hour). A two-stage post-processing pipeline (temporal clustering and FPR-constrained threshold selection) is used to consolidate predictions and generate final submissions.

Key Contributions

Novel Data Representation: The introduction of a channel-based representation for waterfall spectrograms, allowing neural networks to process temporal sequences as input channels similar to color channels in images.
Benchmark Evaluation: A rigorous evaluation of three distinct neural network architectures (MLP, CNN, ViT) on the RADAI benchmark, providing a comparative analysis against the state-of-the-art Non-Negative Matrix Factorization (NMF) baseline.
Performance Analysis: Detailed breakdown of performance across different source categories (Industrial, Medical, NORM, and Nuclear Material), highlighting specific strengths and weaknesses of learned representations versus template-based methods.

Results
The study evaluated the models at three FPR constraint levels. At the constraint of less than one false alarm per hour (FPR < 1):

The CNN outperformed the baseline NMF method across all global metrics, achieving a True Detection Rate (TDR) of 0.4334, True Classification Rate (TCR) of 0.3965, and True Identification Rate (TIDR) of 0.2950. In comparison, NMF achieved 0.4151, 0.3611, and 0.2625, respectively.
The CNN showed particular strength in challenging detection scenarios. For NORM sources (which closely resemble background), the CNN achieved a TDR of 0.0965 compared to 0.0044 for NMF (a >20-fold improvement). For Nuclear Material, the CNN achieved a TIDR of 0.3124 compared to 0.2118 for NMF.
However, at lower FPR constraints (e.g., FPR < 0.25 and FPR < 0.125), the baseline NMF method demonstrated superior performance, suggesting that template-based approaches remain more robust when false alarms must be minimized.
For Industrial sources (with distinctive photopeaks like $^{137}$ Cs and $^{60}$ Co), NMF maintained the highest identification rates, indicating that neural networks may be learning unnecessary complexity for sources with unambiguous spectral signatures.

Significance and Claims
The paper claims that machine learning methods, specifically those utilizing computer vision architectures on waterfall spectrograms, are competitive with and can outperform traditional template-based approaches (like NMF) in mobile urban search scenarios, particularly when moderate false alarm rates are acceptable. The results demonstrate that learned convolutional features can capture subtle spectral patterns and temporal dynamics that fixed templates miss, offering significant advantages for detecting challenging sources like NORM and Nuclear Materials.

However, the authors maintain a modest tone regarding the limitations. They acknowledge that at stricter false positive constraints, neural networks currently underperform compared to NMF. The paper suggests that the optimal strategy may depend on the operational context, with template-based methods excelling for distinctive signatures and learned methods for subtle, complex patterns. The authors propose that future work should explore hybrid approaches, ensemble methods, or two-stage detection/classification pipelines to leverage the complementary strengths of both paradigms.