TLDR: This paper provides the most extensive survey and benchmark of nearly sixty pulse shape discrimination (PSD) algorithms for radiation detection. It classifies methods into statistical and prior-knowledge paradigms, evaluating them on two standardized datasets using various metrics. The study finds that deep learning models, particularly Multi-Layer Perceptrons (MLPs) and hybrid approaches, often outperform traditional methods. It also discusses architectural suitability, limitations of traditional metrics, and the impact of energy thresholds. An open-source Python and MATLAB toolbox with datasets is released to promote reproducibility and advance PSD research.
In the critical field of radiation detection, distinguishing between different types of particles, such as neutrons and gamma-rays, is paramount. This process, known as Pulse Shape Discrimination (PSD), is vital for applications ranging from nuclear reactor monitoring to medical radiopharmaceuticals and fundamental physics research. A recent comprehensive study delves deep into nearly sixty PSD algorithms, offering an unprecedented benchmark and survey of their effectiveness.
The core principle of PSD lies in the unique way different particles interact with a detector medium. When a particle hits a scintillator detector, it generates an electrical pulse. Gamma-rays, interacting with electrons, produce pulses with a rapid decay. Neutrons, on the other hand, interact with protons, leading to a high-density ionization track that results in pulses with a noticeably prolonged tail. By analyzing these distinct temporal characteristics, PSD techniques can accurately identify the incident particle type.
Classifying the Algorithms
The research categorizes PSD algorithms into two main paradigms: statistical discrimination and prior-knowledge discrimination. Statistical methods analyze collected pulses to calculate a discrimination factor, often revealing distinct statistical distributions for different particle types. This category includes time-domain methods like Charge Comparison (CC) and Pulse Gradient Analysis (PGA), and frequency-domain techniques such as Discrete Fourier Transform (DFT) and Wavelet Transform. More recently, neural network-based approaches like Pulse-Coupled Neural Networks (PCNN) have also emerged within this statistical framework.
Prior-knowledge discrimination methods, in contrast, use pre-labeled training datasets to build predictive models. These models learn the distinguishing features of pulse shapes and can then classify new, individual pulses. This paradigm encompasses classical machine learning methods like Support Vector Machines (SVM) and K-Nearest Neighbors (KNN), as well as modern deep learning models including Multi-Layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and advanced architectures like Transformers and Mamba.
The Comprehensive Benchmark
To provide a rigorous comparison, all algorithms were implemented and evaluated on two standardized datasets. The first, from a 241Am-9Be source, represented a typical unlabeled scenario where ground-truth labels are unavailable. For this, the Figure of Merit (FOM) was used to quantify the separation between neutron and gamma-ray distributions. The second dataset, from a 238Pu-9Be source, offered time-of-flight labeled data, enabling direct evaluation using classification metrics like F1-score, accuracy, precision, recall, and ROC-AUC.
A significant finding was that deep learning models, particularly MLPs and hybrid approaches combining statistical features with neural regression, frequently outperformed traditional methods. For instance, the Gated Recurrent Unit (GRU) model, when trained on certain statistical factors, achieved an exceptionally high FOM, showcasing the power of recurrent architectures in capturing temporal dependencies. Similarly, MLPs demonstrated superior F1-scores in direct classification tasks.
Surprising Insights: Simplicity Often Wins
One of the most intriguing revelations from the study is that simpler deep learning models, like Multi-Layer Perceptrons (MLPs), often excel in PSD tasks, sometimes even surpassing the performance of the statistical methods they were trained on. This phenomenon, dubbed ‘a student surpassing its teacher,’ occurs because MLPs, with their fully connected layers, can effectively learn the importance of specific, fixed time segments in the pulse waveform where neutron and gamma-ray signals diverge. They take a ‘global view’ of the pulse, directly exploiting subtle, position-specific details crucial for discrimination without the overhead of searching for patterns.
Conversely, more complex architectures like Convolutional Neural Networks (CNNs), Mamba, and Transformers often struggle. CNNs, designed for translation invariance (finding patterns anywhere), are ill-suited for PSD where key features are in fixed temporal locations. This architectural mismatch can lead to missed features or overfitting. Mamba models, optimized for very long sequences, introduce unnecessary complexity for the short pulse waveforms, leading to computational overhead and sensitivity to hyperparameter tuning. Transformers, with their self-attention mechanism, are computationally expensive and inefficient for local, short-sequence analysis, often requiring vast amounts of data and fine-tuning that are not typical in PSD environments.
Beyond the Traditional Metrics
The research also critically examines the traditional Figure of Merit (FOM) metric, highlighting its limitations, especially when ground-truth labels are available. It advocates for a more robust evaluation using metrics like the F1-score, ROC curves, and Area Under the Curve (AUC), which provide a more balanced and comprehensive assessment of a model’s classification accuracy. For scenarios without ground-truth labels, correlation analysis between different PSD methods is proposed as a valuable supplementary criterion, indicating consistency among high-performing techniques.
The study also investigated the impact of energy thresholds on PSD performance. Generally, performance improved as the energy threshold increased from 2 MeV to 4 MeV, reflecting clearer signal characteristics at higher energies. However, a slight decline was observed at the 5 MeV threshold, attributed to mislabeled non-prompt gamma events in the dataset.
Also Read:
- AI’s New Frontier: Designing Materials with Generative Models
- Automating 4D-STEM Preprocessing for Material Characterization
An Open-Source Future for PSD
To foster reproducibility and accelerate advancements in the field, the researchers have made their comprehensive, dual-language (Python and MATLAB) toolbox and the standardized datasets openly available. This resource can be found at https://github.com/HaoranLiu507/PulseShapeDiscrimination. This initiative aims to empower researchers and engineers to perform direct, fair comparisons and select optimal methods for their specific needs.
Looking ahead, future research will focus on developing novel hybrid architectures that effectively fuse statistical features with deep learning models, expanding training datasets, and integrating multi-modal contextual information (like scintillator type or environmental conditions) to further enhance discrimination accuracy. The ultimate goal remains to optimize these advanced techniques for real-time, on-the-fly implementation in practical field applications.
