DeepSpectrum: A Universal and Accurate Spectral Multimodal Analysis Framework for Lung Cancer Diagnosis

Abstract

In recent years, spectroscopic technology has become a key tool in clinical diagnosis. However, the traditional single-modal spectral analysis method often struggles to capture the full biochemical complexity of heterogeneous biological samples, resulting in limited chemical interpretability and diagnostic sensitivity. To address these challenges, a novel multimodal spectroscopic analysis framework called DeepSpectrum is proposed. It combines the Spectral Mamba and bidirectional gated recurrent unit architectures to enable high-precision lung cancer diagnosis. Specifically, the spectral signals are converted into multichannel image modalities using the continuous wavelet transform (CWT) and Gramian angular difference fields (GADF), thereby constructing rich multimodal data sets. Subsequently, the DeepSpectrum framework is trained, validated, and tested on multimodal spectral data sets, including Raman, Fourier transform infrared, UV–vis absorbance, and fluorescence spectra. It achieved accuracy rates of 98.1%, 98.5%, 92.0%, and 89.2% in lung cancer diagnosis, demonstrating broad applicability across spectral types and complex biological samples. Compared with 11 existing single-modal and multimodal methods, DeepSpectrum showed clear advantages across all evaluation metrics. To further explain the model’s decision-making basis, weighted class activation mapping is used to analyze the spectral signals of normal and cancer samples. The model focuses on key regions such as 1127–1375 cm–1 (heme/hemoglobin vibrations) and 1447 cm–1 (lipid metabolism) in Raman spectra, as well as 1000–1140 cm–1 (nucleic acids) in FTIR spectra. These peak values correspond to the evolution process of lung cancer. This not only validates the biochemical interpretability of the DeepSpectrum model but also provides a reliable diagnostic tool for the precise clinical identification of lung cancer.