FEANet: Frequency-Enhanced Attention Network with detail-preserving fusion for pulmonary disease classification.

Background: Pulmonary disease classification from medical images remains challenging due to heterogeneous imaging modalities and complex texture patterns. Most deep learning methods rely on spatial representations and underutilize frequency-domain information. This study aimed to develop and validate a frequency-enhanced deep learning framework for robust pulmonary disease classification across heterogeneous imaging datasets. Methods: We propose a Frequency-Enhanced Attention Network (FEANet) for pulmonary disease classification across multiple imaging modalities. FEANet employs a dual-branch architecture based on discrete wavelet transform to separate low-frequency (LF) structural information and high-frequency (HF) texture details. A Wavelet-Aware Semantic Context Attention (WASCA) module models global semantic information, a Frequency-Guided Detail Attention (FGDA) module adaptively enhances discriminative HF features, and a Frequency-Conditioned Fusion Attention (FCFA) module integrates complementary frequency representations. The model was evaluated on lung cancer computed tomography (CT), coronavirus disease 2019 (COVID-19) chest X-ray, and a clinical COPD CT dataset. Results: FEANet achieved accuracies of 98.70%, 96.04%, and 99.84%, with corresponding area under the curves (AUCs) of 0.9995, 0.9951, and 0.9998 on the lung cancer CT, COVID-19 X-ray, and COPD CT datasets, respectively. On the COVID-19 dataset, the full FEANet model achieved 96.04% accuracy and 96.04% F1-score, outperforming the strongest compared baseline by 0.52 percentage points in accuracy. Ablation analysis showed that the combined use of LF semantic attention, HF detail attention, and frequency-conditioned fusion produced the best performance. Robustness analysis further showed stable performance under mild-to-moderate image perturbations. Conclusions: FEANet effectively integrates frequency- and spatial-domain features, improving robustness and generalization across heterogeneous imaging modalities. It shows potential for clinical computer-aided diagnosis.