Attention Scale Fusion Network for Qualitative and Quantitative Analysis of Serum Tumor Biomarkers Via Label-Free Surface-Enhanced Raman Spectroscopy

Serum tumor biomarkers are critical molecular indicators reflecting tumor initiation and progression, making it essential for developing highly sensitive and convenient detection methods. Label-free surface-enhanced Raman spectroscopy (SERS) is a powerful analytical method, offering detailed molecular fingerprints of biomolecules including tumor biomarkers. However, its application in serum biomarker analysis remains challenging due to matrix interference obscuring spectral features of low-abundance analytes. Here, we developed an attention scale fusion network (ASFN) applied for label-free SERS data to detect and quantify biomarkers in serum. ASFN employs a multiscale dual-branch convolutional architecture, integrates Transformer modules for adaptive feature fusion, and introduces a task-prior transfer mechanism, significantly enhancing the model's robustness. Compared to other single-task machine learning and deep learning methods, the proposed approach demonstrates superior performance and successfully overcomes the limitations of spectral similarity. It achieves 100% classification accuracy and a weighted <i>R</i>² of 0.9713 for concentration prediction in serum sample analysis. Moreover, the application of Permutation Importance visualization provides valuable insights into the decision-making mechanism of ASFN. Hence, this work not only provides a novel technical approach for complex system analysis but also opens up new mind for the application of multitask deep learning in biomedical detection.