可解释性
化学
人工智能
元素分析
模式识别(心理学)
拉曼光谱
判别式
模态(人机交互)
光谱学
解码方法
光谱分析
计算机科学
生物系统
光谱特性
肾细胞癌
激光器
癌症
光谱成像
分析技术
偏最小二乘回归
作者
Zehai Hou,Shengkun Yan,Keyao Li,Furi Ling,Xiong Wang,Lianbo Guo
标识
DOI:10.1021/acs.analchem.6c04864
摘要
Abstract Spectroscopic analysis of biofluid shows strong potential for cancer detection; however, the complexity of biological matrices can reduce the discriminative performance of conventional single modality spectroscopy, and challenges such as limited interpretability remain. Here, we developed a multimodal spectroscopic system that analyzes the same sample by integrating surface-enhanced Raman spectroscopy (SERS) and laser-induced breakdown spectroscopy (LIBS). By leveraging complementary molecular and elemental information from these two modalities, we propose MambaTrans-KAN (MTKNet), a heterogeneous dual-branch spectral fusion network, for the classification of major urological cancer types, including bladder cancer, prostate cancer, and renal cell carcinoma, and for interpretable analysis of disease-associated spectral features. The proposed method achieved an accuracy of 97.25%, improving performance by approximately 14.25% over single modality spectroscopy, with a macro-averaged precision of 98.07% and recall of 97.42%. In addition, we introduce a Kolmogorov-Arnold Network (KAN) for multimodal spectral decoding and use its structure for model visualization, providing interpretable evidence for the decision-making process. This analysis revealed elemental features associated with Ca, Mg, K, and Na, as well as molecular spectral features related to galactose metabolism, phenylalanine associated vibrations, and purine related components, which may reflect disease associated biochemical alterations. By combining noninvasive urine analysis with an interpretable multimodal spectrochemical framework, this strategy provides a potential analytical approach for extracting complementary molecular and elemental information from complex biofluids and improving the understanding of disease associated spectral signatures.
科研通智能强力驱动
Strongly Powered by AbleSci AI