Multimodal diagnostic network integrating infrared and mass spectra for lung cancer.

Lung cancer has high global morbidity and mortality, and accurate early diagnosis is critical for improving prognosis. Fine-needle aspiration (FNA) is a minimally invasive initial screening tool, but traditional analytical methods (e.g., hematoxylin and eosin staining) are subjective and prone to misdiagnosis. Fourier-transform infrared (FTIR) spectroscopy enables label-free analysis of biomolecular vibrations, and mass spectrometry detects metabolite changes. Here, FTIR and mass spectra were collected from FNA samples of lung cancer, with histopathology adopted as the diagnostic gold standard. A multimodal diagnostic network was developed, comprising modality-specific feature extraction branches and a hybrid fusion module, which integrates gated fusion and multi-head cross-attention mechanisms to capture correlations between FTIR and mass spectra. This architecture achieved an area under the curve (AUC) of 96.67% and a precision of 91.42%, outperforming single-modal approaches and conventional methods. Furthermore, model interpretability analysis was performed, which tentatively identified potential key biomarkers for lung cancer. This work seeks to provide a minimally invasive, rapid, and accurate tool for lung cancer diagnosis, thereby facilitating early clinical intervention and improving patient prognosis.