ABSTRACT Early diagnosis of cancer is crucial for improving survival rates and treatment outcomes. However, conventional diagnostic methods, such as tissue biopsy and imaging, often involve invasive procedures and may require extended processing times. In recent years, noninvasive biomarker detection techniques, such as Raman spectroscopy and infrared (IR) spectroscopy, have been widely applied in cancer diagnosis. Nevertheless, spectral data obtained from different modalities are frequently affected by factors such as sampling frequency, noise, and acquisition conditions, making data alignment and fusion increasingly challenging. To enhance the accuracy of multi‐class cancer biomarker identification, this paper proposes a multimodal intelligent recognition model based on the combined analysis of Raman and IR spectroscopy. The model first aligns the spectral data from different modalities using the synchronous dynamic time warping (SDTW) algorithm to eliminate frequency shifts. Then, we apply fast Fourier transform (FFT) to extract frequency–domain features, capturing key information from the spectra. Finally, a dynamic convolution module (DynamicConv1d) is introduced, which adaptively generates convolutional kernels through a learning mechanism. This allows for customized feature extraction tailored to different spectral samples, effectively enhancing feature discriminability. The model fuses information from both average pooling and max pooling to highlight important channels related to molecular vibrations while suppressing noise interference. In the feature fusion stage, a dynamic feature fusion module is constructed. It employs an attention mechanism to perform a weighted integration of Raman and IR features, fully leveraging the complementarity between the two modalities. Experimental results demonstrate that the proposed method achieves superior multi‐class performance on several cancer spectral datasets. It can not only diagnose the presence of cancer but also differentiate between specific types, significantly improving the model's generalization ability and diagnostic reliability. This study provides new ideas and methodological support for the application of multimodal spectroscopic techniques in early disease screening and biomarker detection. The experimental results show that the model achieves an accuracy of 98.55% and an AUC of 0.9900. It also outperforms traditional single‐modality models across multiple evaluation metrics and has strong potential for clinical application.
Li et al. (Fri,) studied this question.