Chen Xiangnan, Zhou Xuguang, Lv Xiaoyi, Wu Lijun, Li Jiahe, Chen Chen, Chen Cheng
College of Software, Xinjiang University, Urumqi 830046, China.
Department of Rheumatology and Immunology, People's Hospital of Xinjiang Uyghur Autonomous Region, Xinjiang, China.
Spectrochim Acta A Mol Biomol Spectrosc. 2026 Jan 15;345:126836. doi: 10.1016/j.saa.2025.126836. Epub 2025 Aug 18.
In today's medical diagnosis field, accurate diagnosis of many diseases plays an important role. As emerging non-invasive diagnostic technologies, Raman spectroscopy and infrared spectroscopy have shown unique advantages in the detection of disease characteristic markers due to their advantages such as high sensitivity and specificity. However, the diagnosis of some diseases by using only Raman or infrared single-spectral technology is relatively insufficient. Therefore, how to effectively utilize the two spectral data and accurately diagnose the disease as early as possible has become an important challenge.At present, the fusion of two spectral data types has shown good performance in disease diagnosis. Compared with single-spectral technology, multimodal fusion can more comprehensively reflect disease-related information, thereby improving the accuracy and reliability of diagnosis. However, existing research is usually limited to a single disease, and the inter-modal feature conflicts in the multimodal fusion process have not been effectively addressed. The conflict and overlapping information between the two in certain frequency bands and features are often ignored, and the fusion is mainly focused on shallow levels.Based on the above problems, this paper proposes a disease diagnosis technology based on the fusion of multi-spectral matching and a collaborative attention mechanism of Raman and infrared spectroscopy. Through the frequency band adaptive decomposition mechanism combined with a dynamic weighting strategy of multi-spectral attention, the problems of modality conflict and redundant fusion in traditional methods are effectively solved. Cross-modal deep interaction is then achieved through residual feature injection, allowing for cross-modal complementary enhancement of key features for early diagnosis while maintaining modality specificity.Comprehensive experimental results show that the model has excellent performance in the diagnosis of thyroid cancer and systemic lupus erythematosus(SLE), reaching 98.46 % and 98.57 % accuracy, 98.46 % and 99.00 % recall, and 98.93 % and 98.50 % AUC values, respectively, providing a new method and idea for disease diagnosis using multimodal spectral data.
在当今医学诊断领域，许多疾病的准确诊断起着重要作用。作为新兴的非侵入性诊断技术，拉曼光谱和红外光谱由于其高灵敏度和特异性等优点，在疾病特征标志物检测中显示出独特优势。然而，仅使用拉曼或红外单光谱技术对某些疾病进行诊断相对不足。因此，如何有效利用这两种光谱数据并尽早准确诊断疾病已成为一项重要挑战。目前，两种光谱数据类型的融合在疾病诊断中已显示出良好性能。与单光谱技术相比，多模态融合能够更全面地反映疾病相关信息，从而提高诊断的准确性和可靠性。然而，现有研究通常局限于单一疾病，多模态融合过程中的模态间特征冲突尚未得到有效解决。两者在某些频段和特征上的冲突和重叠信息常被忽视，融合主要集中在浅层。基于上述问题，本文提出一种基于拉曼光谱和红外光谱多光谱匹配与协同注意力机制融合的疾病诊断技术。通过结合多光谱注意力动态加权策略的频带自适应分解机制，有效解决了传统方法中的模态冲突和冗余融合问题。然后通过残差特征注入实现跨模态深度交互，在保持模态特异性的同时，实现关键特征的跨模态互补增强以进行早期诊断。综合实验结果表明，该模型在甲状腺癌和系统性红斑狼疮（SLE）诊断中具有优异性能，准确率分别达到98.46%和98.57%，召回率分别达到98.46%和99.00%，AUC值分别达到98.93%和98.50%，为利用多模态光谱数据进行疾病诊断提供了一种新方法和新思路。
