Kan-AAE-driven synthetic SERS spectra generation method for Precise cancer identification.

Surface-Enhanced Raman Spectroscopy (SERS) is gaining popularity in cancer detection studies because it offers a non-invasive and rapid approach. Label-free SERS detection techniques often needs machine learning, which depends on adequate data for training. The scarcity of blood serum samples from cancer patients, due to challenges in collection linked to confidentiality concerns and other restrictions, can result in model overfitting and poor generalization ability. To tackle this challenge, we propose the KAN-AAE method, a new approach for creating synthetic SERS spectra, which lever- ages the power of Kolmogorov-Arnold Networks (KAN) in conjunction with Adversarial Autoencoders (AAE) and has the excellent capability of fitting the distribution of complex data in feature space. We conducted experiments by collecting serum samples from patients with four different types of cancer, two types of other diseases, and healthy individuals, subsequently measuring their SERS spectra. We trained the KAN-AAE model using the SERS spectral data and used it to produce synthetic spectra, which were then combined with actual data for classifier training, enhancing data diversity. Utilizing the combined dataset, there was a notable increase of 1% to 3% in the accuracy of classification models like logistic regression, decision tree, multilayer perceptron, 1D-convolutional neural network, and KAN. The KAN classifier outperformed others, achieving an accuracy rate of 95.62%. The experimental results demonstrate that: (1) our proposed method gener- ates high-quality and reliable SERS spectra data; (2) the method effectively improves the classification accuracy for various types of cancer.