EfficientMFRF-BiFPN: A CNN-Transformer Hybrid With Multi-Feature Representation Fusion for Pneumonia Detection

Abstract

Pneumonia remains a significant global health threat, especially in vulnerable populations, and early, accurate diagnosis is critical for effective intervention. While deep learning has significantly advanced medical image analysis, most existing architectures remain limited in their ability to jointly model local lesion features and global semantic context. To address this, we propose EfficientMFRF-BiFPN, a novel CNN–Transformer hybrid architecture that introduces two key innovations: a Multi-Feature Representation Fusion (MFRF) module and an improved Bidirectional Feature Pyramid Network (BiFPN). The MFRF module effectively reconciles and integrates heterogeneous CNN and Transformer features within a hybrid framework. This module utilizes bidirectional feature projection and cross-attention to enable dynamic, context-aware interaction between local and global representations. The subsequent convolutional feature enhancement and gated fusion mechanisms adaptively balance both intra-branch and inter-branch information flows, further enhancing the overall feature representation capability. In parallel, the improved BiFPN integrates CNN and Transformer features, enhancing heterogeneous feature interactions and balancing semantic and spatial information to generate rich multi-scale feature representations. Together, these components form a three-level fusion paradigm—local, global, and multiscale—that enables the model to comprehensively capture diverse feature representations across spatial and semantic hierarchies, significantly improving pneumonia classification performance on chest X-ray images. On a public chest X-ray dataset, our model achieves excellent results (accuracy: 97.77%, recall: 98.8%, F1-score: 98.49%), demonstrating high diagnostic reliability and strong potential as a resource-efficient tool for clinical pneumonia screening.