PneuSwin: An Advanced X-Ray-Based Diagnostic System Integrating Ensemble Deep Learning Architectures and Swin Transformers for Pneumonia Detection

Abstract

Pneumonia continues to pose a considerable global health concern, characterized by elevated fatality rates globally. Xrays are the primary radiological imaging technique for detecting pneumonia because of their widespread availability and inexpensive cost in medicine. Researchers have employed a variety of Deep Learning (DL)-based procedures to solve the issue, but only a small number of studies have amalgamated DL methods with Swin transformers. The Swin Transformer, distinguished for its ability to capture long-range dependencies and spatial associations, subsequently processes the refined features. This investigation introduces PneuSwin, a novel X-ray-based diagnostic system for efficient detection. The study utilized the PneuData dataset, a composite of three public X-ray datasets. It had Bacterial Pneumonia (BP), Viral Pneumonia (VP), and normal classes with balanced instances. Initially, the study paired various DL architectures (CapsNet, DenseNet- 121, EfficientNet-B3, and ResNet-101) with the swin transformer. The DLs extracted relevant features from the PneuData images, sent them to Principal Component Analysis (PCA) to diminish their dimension and pick out the relevant features, and then inserted them as patches into the swin transformer for either binary or multi-class classification. Conversely, the PneuSwin concatenates the retained features, transferring them as a feature matrix to the PCA for relevant features and feeding them to the swin transformer. In both binary and multi-class classification, PnewSwin outperformed in comparison to ensemble DLs. In binary, PneuSwin had 97.21% accuracy, a 96.95% F1 score, and a 0.967 AUC, while in multi-class it had 97.67% accuracy, a 97.31% F1 score, and a 0.973 AUC. The results indicate that PneuSwin is proficient in detecting pneumonia in both binary and multi-class classifications.