Optimizing AI Models for Raman Spectroscopy: Improving Disease Diagnosis

A new study published in Analytical Chemistry demonstrated how the optimization of artificial intelligence (AI) models using Raman spectroscopy has improved disease diagnosis (1).

Raman spectroscopy is a powerful tool for the label-free biomolecular analysis of cells and tissues, utilized extensively in both in vitro and in vivo pathological diagnosis (1). The integration of this molecular spectroscopic technique with AI and machine learning has helped create new ways to improve disease diagnosis, such as Alzheimer’s (1,2). Some of these machine learning algorithms include principal component analysis–support vector machine (PCA-SVM) and SVM, as well as manifold learning techniques like uniform manifold approximation and projection (UMAP) and deep learning models including ResNet and AlexNet. However, the challenge has been to determine the optimal AI classification model for various Raman spectral data sets with differing characteristics.

Abstract futuristic digital and technology on dark blue color background. AI(Artificial Intelligence) wording with the circuit design. | Image Credit: © Thitichaya - stock.adobe.com

Researchers from Capital Medical University and Beihang University explored this topic. The study was led by Limin Feng and Shuhua Yue, and it focused on how they could use different AI technologies to improve the accuracy of pathological diagnoses by analyzing different types of Raman spectral data (1).

In their study, research team chose five representative Raman spectral data sets: endometrial carcinoma, hepatoma extracellular vesicles, bacteria, melanoma cells, and diabetic skin. These data sets varied significantly in terms of sample size, spectral data size, Raman shift range, tissue sites, Kullback–Leibler (KL) divergence, and significant Raman shift positions (wavenumbers showing notable differences between groups) (1). The goal of the study was to evaluate the performance of different AI models, such as PCA–SVM, SVM, UMAP–SVM, Residual Network (ResNet), and Alex Network (AlexNet), adjusting the network parameters based on these data characteristics (1).

For data sets with large spectral data sizes, the deep learning model ResNet outperformed PCA-SVM and UMAP-SVM. The performance of ResNet indicated that deep learning models might be more suitable for handling extensive spectral data because of their ability to manage complex data patterns and large volumes of information (1). By building a data characteristic-assisted AI classification model, the researchers were able to fine-tune network parameters such as the number of principal components, activation functions, and loss functions based on data size and KL divergence, among other factors (1).

By tinkering with the network parameters, the research team demonstrated improvement in the accuracy of AI models. The AI models used for diagnosing endometrial carcinoma improved from 85.1% to 94.6% (1). Similarly, the accuracy for detecting hepatoma extracellular vesicles increased from 77.1% to 90.7%, while melanoma cell detection accuracy rose from 89.3% to 99.7% (1). For bacterial identification, the accuracy jumped from 88.1% to 97.9%, and diabetic skin screening saw a significant improvement from 53.7% to 85.5% (1). Furthermore, the mean time expense for these analyses was reduced to approximately 5 s, demonstrating the efficiency and speed of the optimized AI models.

These advancements not only highlight the potential of AI in enhancing the diagnostic capabilities of Raman spectroscopy but also underscore the importance of tailoring AI models to the specific characteristics of the spectral data. The study by Feng, Yue, and their team paves the way for more accurate, efficient, and reliable pathological diagnoses, which could have profound implications for medical practice and patient outcomes (1).

By leveraging the strengths of various AI technologies and optimizing them for specific data characteristics, the researchers have demonstrated that significant improvements in diagnostic accuracy and efficiency are achievable using these tools in conjunction with Raman spectroscopy.

References

(1) Chen, X.; Shen, J.; Liu, C.; et al. Applications of Data Characteristic AI-Assisted Raman Spectroscopy in Pathological Classification. Anal. Chem. 2024, 96 (16), 6158–6169. DOI: 10.1021/acs.analchem.3c04930

(2) Wetzel, W. Applying Raman Spectroscopy with Machine Learning to Detect Alzheimer’s Disease. Spectroscopy. Available at: https://www.spectroscopyonline.com/view/applying-raman-spectroscopy-with-machine-learning-to-detect-alzheimer-s-disease (accessed 2024-06-19).