Neutron/Gamma Pulse Shape Discrimination via Scalogram Imaging and Pretrained CNN

Neutron/gamma Pulse Shape Discrimination (PSD) is a critical task in radiation detection systems, where reliable classification becomes increasingly challenging under low Signal-to-Noise Ratio (SNR) conditions. Although Deep Convolutional Neural Networks (DCNNs) have demonstrated promising performance in automated signal classification, their robustness can degrade when detector signals are affected by varying noise levels. This paper proposes a Physics-Guided Feature Fusion with AlexNet (PGFF-AlexNet) framework that integrates time-frequency representations learned from scintillation pulses with complementary domain-specific physical descriptors. First, one-dimensional neutron and gamma pulses are transformed into two-dimensional Continuous Wavelet Transform (CWT) scalogram images to preserve their temporal and spectral characteristics. A comparative evaluation of four pretrained Convolutional Neural Network (CNN) architectures (AlexNet, GoogLeNet, EfficientNet-B0, and MobileNetV2) is then conducted over SNR levels ranging from 5 to 35 dB to identify a suitable visual feature extractor. Based on its overall discrimination capability and computational characteristics, AlexNet is selected as the visual backbone of the proposed framework. Its deep scalogram features are fused with a low-dimensional physics branch incorporating charge-comparison, frequency-domain, rise-time, and pulse-gradient descriptors. The resulting visual and physics-based representations are projected and concatenated prior to classification, enabling the network to exploit both learned time-frequency patterns and physically meaningful pulse-shape information. Experimental results show that PGFF-AlexNet achieves accuracies of 0.7369, 0.8900, and 0.9644 at 5, 10, and 15 dB, respectively, and an aggregate multi-SNR accuracy of 0.9548. The proposed framework provides substantial improvements over the standalone CNN baselines under challenging low-to-moderate SNR conditions and achieves an aggregate Matthews Correlation Coefficient (MCC) of 0.9090 and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.9912. Feature-space analyses using Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and intermediate activation maps further illustrate the complementary contribution of the physics-guided representation. These results demonstrate that PGFF-AlexNet provides an accurate, noise-robust, and interpretable framework for neutron/gamma PSD across diverse operating conditions.