A Dual-Structured Stacked ELM-AE Network for Multi-Class Chest X-Ray Image Classification

This paper investigates the application of a dual-structured stacked extreme learning machine autoencoder to multi-class chest X-ray image classification, a task often constrained by the limited contrast, high noise levels, and sparse features of X-ray images. The approach employs multiple stacked ELM-AE modules for layer-wise feature extraction and compression, using orthogonally initialized input weights to ensure uncorrelated feature mapping and improved numerical stability. Bayesian optimisation is incorporated to adaptively determine the number of hidden nodes and the regularization parameter, reducing manual tuning and enhancing generalization. Features of 2048 dimensions are extracted from the avg-pool layer of a pre-trained ResNet-50, and two output strategies—a direct classifier and an additional ELM classifier—are evaluated under identical feature conditions. Experimental results on the COVID-19 Radiography Database demonstrate that the two-layer DELM-AE with a direct classifier achieves a balanced trade-off between accuracy, generalization, and computational efficiency, attaining 90.69% ± 0.16 test accuracy while significantly reducing runtime compared with support vector machine and random forest classifiers. The method provides stable performance with low training overhead, highlighting its potential for efficient and interpretable chest X-ray image classification in resource-constrained clinical settings.