Automated threat detection in X-ray security imagery is a critical yet challenging task, where conventional deep learning models often struggle with low accuracy and overfitting. This study addresses these limitations by introducing a novel framework based on feature fusion. The proposed method extracts features from multiple and diverse deep learning architectures and classifies them using a Random Weight Network (RWN), whose hyperparameters are optimized for maximum performance. The results show substantial improvements at each stage: while the best standalone deep learning model achieved a test accuracy of 83.55%, applying the RWN to a single feature set increased accuracy to 94.82%. Notably, the proposed feature fusion framework achieved a state-of-the-art test accuracy of 97.44%. These findings demonstrate that a modular approach combining multi-model feature fusion with an efficient classifier is a highly effective strategy for improving the accuracy and generalization capability of automated threat detection systems.

Human intelligence has historically produced various solutions to the challenges it has faced thanks to its cognitive abilities such as learning, remembering, comparing, and problem-solving. One of the most fundamental examples of artificial intelligence approaches that emerged by modeling these behaviors specific to intelligence is the artificial neural network (ANN), inspired by the biological nervous system. ANNs, with their similarity to biological nervous systems, present a highly complex system, and with technological advancements, the modeling and applicability of these systems are constantly increasing. In this study, a new ANN architecture inspired by the biological nervous system was designed. To improve the classification performance of the developed architecture, the connections in the network were treated as a binary optimization problem and optimized using the Binary Particle Swarm Optimization (BPSO) algorithm. Following the architecture design, the weights of the connections in the network were optimized using the Particle Swarm Optimization (PSO) algorithm which works with continuous values. As a result, a network architecture compatible with the biological nervous system, flexible in structure, and highly applicable to various problem types was created. The developed method was tested on Breast Cancer, Hepatitis, Mammographic Mass and Haberman’s Survival datasets. Additionally, proposed model compared with the Genetic Algorithm (GA), one of the most widely used evolutionary computation methods in the literature for connection optimization. Specifically, the proposed model reduces the input layer size from 5 to 2 in the Mammographic Mass domain by applying aggressive feature selection. It successfully reduces computational complexity by eliminating up to 12 hidden neurons in the Hepatitis dataset without sacrificing classification accuracy, thus providing significant structural pruning. Furthermore, it demonstrates high stability across all metrics with minimal standard deviation, proving that the synergy between binary and continuous swarm variants generates highly concise, lightweight, and efficient neural network topologies