Pulse Shape Discrimination Algorithms: Survey and Benchmark

This paper presents a comprehensive survey and benchmark of nearly sixty pulse shape discrimination algorithms, demonstrating through standardized evaluation on two datasets that deep learning models often outperform traditional methods while providing an open-source toolbox to advance reproducibility in radiation detection research.

The Gist

Imagine you are a bouncer at a very exclusive, high-security club. The club is a radiation detector, and the guests are tiny particles flying through the air. The problem? Two very different types of guests are trying to get in: Neutrons (the VIPs you want to let in) and Gamma Rays (the imposters you need to keep out).

Both guests look almost identical when they first walk through the door. They both wear the same "fast-rising" jacket. However, as they move deeper into the club, their behavior changes. The VIPs (Neutrons) are slow to leave; they linger in the hallway, dragging their feet and taking a long time to exit. The imposters (Gamma Rays) are energetic and sprint out immediately.

Pulse Shape Discrimination (PSD) is the art of watching how long these guests linger to tell them apart. For decades, scientists have built different "rules" and "algorithms" to make this judgment.

This paper is like a massive, head-to-head Olympic competition where the authors put nearly 60 different bouncer strategies against each other to see who is the best at spotting the imposters.

The Three Teams of Bouncers

The authors organized the competitors into three main teams:

1. The Statisticians (The Old School): These bouncers use simple math rules. They might measure exactly how much time the guest spends in the hallway versus the total time they were there.
   • Analogy: Imagine a bouncer with a stopwatch and a calculator. "If the guest stays longer than 50% of the total time, they are a VIP." It's fast, simple, and doesn't need to learn anything beforehand.
2. The Machine Learners (The Students): These bouncers are given a photo album of past guests (labeled as "VIP" or "Imposter") and asked to find patterns. They might look at the shape of the guest's walk.
   • Analogy: A bouncer who has seen thousands of photos and says, "I've seen this specific type of slow shuffle before; that's a VIP."
3. The Deep Learning Experts (The Geniuses): These are the most advanced AI models. They don't just look at one feature; they analyze the entire "movie" of the guest's movement, layer by layer, to find hidden clues.
   • Analogy: A bouncer with a supercomputer brain that can analyze the guest's gait, the way they hold their coat, and the rhythm of their footsteps all at once to make a perfect guess.

The Big Surprise: The "Simple" Geniuses Won

You might expect the most complex, high-tech AI (like the "Deep Learning" team) to crush the competition. And while they did very well, the paper found a twist that surprised everyone:

The Simple Multi-Layer Perceptron (MLP) was the MVP.

Think of an MLP as a very smart, but structurally simple, bouncer. It doesn't try to be fancy. It just looks at the specific moment in time where the VIPs and imposters act differently (the "tail" of the signal) and assigns a weight to it.

• Why it won: The paper explains that radiation pulses are short and specific. You don't need a complex camera that scans the whole room (like a Convolutional Neural Network or CNN) to find a specific person in a crowd. You just need to look at the exact spot where the VIP is standing. The simple MLP is perfectly built for this "spot-check" job, while the fancy, complex models often get confused by trying to look everywhere at once.

The "Student Can Beat the Teacher" Trick

One of the coolest findings is a "Hybrid" strategy.
Imagine you have a mediocre teacher (a simple statistical rule) who is good at spotting VIPs but makes mistakes. You then hire a super-smart AI student to watch that teacher.

• The Result: The student doesn't just copy the teacher; the student learns from the teacher but then uses its own brain to fix the teacher's mistakes. In many cases, the student ended up being better than the teacher, achieving higher accuracy than the original rule ever could.

The Scoreboard: How Did They Measure?

The authors didn't just guess who won; they used a scoreboard with four different metrics:

1. FOM (Figure of Merit): The old-school score. It checks how clearly the two groups separate on a graph.
2. F1-Score: The modern, strict score. It checks how many VIPs you actually caught versus how many imposters you accidentally let in.
3. ROC-AUC: A curve that shows how good the bouncer is at any level of strictness.
4. Correlation: They checked if the winners agreed with each other. If two different methods both say "That's a VIP," they are likely right.

The Verdict: The old-school "Figure of Merit" (FOM) isn't always the best judge. Sometimes a method looks great on a graph but actually lets in too many imposters. The F1-Score and ROC-AUC were found to be the true champions of accuracy.

The Takeaway for Everyone

1. Don't Overcomplicate: You don't always need the most expensive, complex AI to solve a problem. Sometimes, a simple, well-designed tool (like the MLP) is the best because it fits the job perfectly.
2. Hybrid is King: Combining a simple, fast rule with a smart AI learner often gives you the best of both worlds: speed and high accuracy.
3. Open Source: The authors didn't just write a paper; they built a public toolbox (like a free app store for scientists) containing all these 60 methods. They want everyone to be able to test these ideas and improve radiation detection for nuclear safety, medical imaging, and space exploration.

In a nutshell: This paper is a massive "taste test" of 60 different ways to tell radiation particles apart. It proves that while fancy AI is powerful, sometimes the simplest, most direct approach is the champion, and the best results come from letting a smart AI learn from a simple rule.

Technical Summary

1. Problem Statement

Pulse Shape Discrimination (PSD) is a critical technique in radiation detection used to distinguish between different types of incident particles (e.g., neutrons vs. gamma-rays) within mixed radiation fields. While numerous algorithms have been proposed over the decades—ranging from traditional statistical methods to modern deep learning (DL) models—the field lacks a unified, systematic, and quantitative benchmark.

• The Gap: Existing literature often evaluates algorithms on disparate datasets with varying metrics, making direct, fair comparisons impossible.
• The Challenge: Researchers struggle to select the optimal method for specific applications due to the sheer volume of diverse approaches and the lack of standardized evaluation protocols.

2. Methodology

The authors conducted a comprehensive survey and benchmark of nearly 60 distinct PSD algorithms, classifying them into two main paradigms:

A. Algorithm Taxonomy

1. Statistical Discrimination (Unsupervised/Rule-based):
   • Time-Domain: Charge Comparison (CC), Charge Integration (CI), Falling-Edge Percentage Slope (FEPS), Gatti Parameter (GP), Log-Likelihood Ratio (LLR), Log Mean Time (LMT), PCA, Pulse Gradient Analysis (PGA), Pattern Recognition (PR), Zero-Crossing (ZC).
   • Frequency-Domain: Discrete Fourier Transform (DFT), Frequency Gradient Analysis (FGA), Fractal Spectrum (FS), Scalogram-based Discrimination (SD), Simplified Digital Charge Collection (SDCC), Wavelet Transforms (Haar, Marr).
   • Neural Network-based (Statistical): Ladder Gradient (LG), Pulse-Coupled Neural Network (PCNN), Random-Coupled Neural Network (RCNN), Spiking Cortical Model (SCM).
2. Prior-Knowledge Discrimination (Supervised):
   • Machine Learning (ML): Boosted Decision Tree (BDT), Decision Tree (DT), Fuzzy C-Means (FCM), Gaussian Mixture Model (GMM), K-Nearest Neighbors (KNN), Linear/Logistic Regression, Learning Vector Quantization (LVQ), Support Vector Machine (SVM), Tempotron.
   • Deep Learning (DL):
     • CNNs: 1D-CNN (Shallow), 2D-CNN (Deep), CNN with Snapshot Plotting, CNN with Fourier/STFT/Wavelet features.
     • MLPs: Various depths (1 to 7 hidden layers) with raw or transformed inputs (FFT, PCA, STFT, Wavelet).
     • RNNs: Simple RNN, Elman, LSTM, GRU.
     • Transformers & Mamba: Self-attention mechanisms and State-Space Models (SSM).

B. Experimental Setup & Datasets

The study utilized two standardized datasets to ensure rigorous evaluation:

1. $^{241}$Am-9Be Dataset (Unlabeled): ~9,414 pulses from an EJ299-33 plastic scintillator. Used to evaluate statistical methods via Figure of Merit (FOM) and to train DL/ML models as regressors (predicting statistical PSD factors).
2. $^{238}$Pu-9Be Dataset (Labeled): ~21,001 pulses from YAP:Ce and NE213A detectors with Time-of-Flight (TOF) ground truth. Used for direct classification evaluation using Accuracy, Precision, Recall, F1-score, and ROC-AUC.

C. Evaluation Metrics

• FOM: Traditional metric for histogram separation (used on unlabeled data).
• F1-score, Precision, Recall, Accuracy: For classification performance on labeled data.
• ROC-AUC: To evaluate discrimination capability across all thresholds.
• Correlation Analysis: Pearson correlation between PSD factors of different methods to assess consistency.
• Computational Cost: Parameters, FLOPs, and inference time.

3. Key Contributions

1. Comprehensive Benchmark: The first large-scale, systematic comparison of ~60 PSD algorithms on unified datasets.
2. Open-Source Toolbox: Release of a dual-language (Python/MATLAB) toolbox containing implementations of all surveyed algorithms and the datasets to promote reproducibility.
3. Hybrid Approach Validation: Demonstrated that DL models trained as regressors on statistical features can outperform the statistical "teachers" themselves.
4. Architectural Insights: Provided a critical analysis of why certain architectures (MLPs) excel while others (CNNs, Transformers, Mamba) struggle with PSD data characteristics.
5. Metric Evolution: Argued for the prioritization of F1-score and ROC-AUC over FOM, especially when ground-truth labels are available, as FOM can be misleading regarding actual classification accuracy.

4. Key Results

• Performance Hierarchy:
  • Deep Learning (DL) generally outperformed traditional statistical methods.
  • MLPs (Multi-Layer Perceptrons) emerged as the top performers, particularly when combined with feature extraction (e.g., MLP2STFT achieved an F1-score of 0.962).
  • Hybrid Models: DL regressors trained on high-performing statistical features (like RCNN or CI) achieved the best results. For instance, GRU-CI achieved an FOM of 1.85, significantly higher than statistical baselines.
  • Underperformers: Complex architectures like Transformers and Mamba underperformed, often failing to capture the specific local features required for PSD. CNNs also struggled, likely due to translation invariance being a liability for fixed-temporal features in pulses.
• Energy Threshold Impact: Most methods showed improved performance as energy thresholds increased from 2 MeV to 4 MeV, with a slight decline at 5 MeV due to mislabeled non-prompt gamma events. Top methods (e.g., MLP1STFT, CI) remained robust across thresholds.
• Efficiency vs. Accuracy: While statistical methods are traditionally faster, optimized DL models (like MLPs and simple RNNs) on modern GPUs (RTX 4090) achieved inference times comparable to or faster than complex statistical methods (e.g., Wavelet transforms), narrowing the efficiency gap.
• Feature Extraction: Charge Integration (CI) and neural-based statistical methods (PCNN, SCM, RCNN) were identified as the most effective "teachers" for training DL regressors.

5. Significance and Implications

• Paradigm Shift: The study challenges the notion that complex, state-of-the-art architectures (Transformers, Mamba) are always superior. It highlights that architectural simplicity (MLPs) is often better suited for the specific, position-dependent nature of pulse data.
• Hybrid Future: The most effective strategy identified is a hybrid approach: using robust statistical methods to extract initial features, which are then refined by lightweight DL regressors. This combines the interpretability and efficiency of statistical methods with the pattern-recognition power of DL.
• Standardization: By providing open datasets and code, the authors establish a new standard for evaluating PSD algorithms, moving the field away from fragmented, non-comparable studies toward a unified, data-driven science.
• Practical Application: The findings offer clear guidance for engineers in nuclear physics, reactor monitoring, and radiopharmaceuticals to select the most efficient and accurate PSD algorithms for real-world deployment.