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

Area of Science:

• Biomedical Spectroscopy
• Machine Learning in Oncology
• Computational Biology

Background:

• Surface-Enhanced Raman Spectroscopy (SERS) offers a non-invasive, rapid method for cancer detection.
• Label-free SERS requires machine learning, which is hindered by limited patient data, leading to overfitting and poor generalization.
• Scarcity of blood serum samples is a significant challenge due to collection complexities and confidentiality concerns.

Purpose of the Study:

• To address the data scarcity issue in SERS-based cancer detection.
• To propose and validate a novel method for generating synthetic SERS spectra.
• To enhance the performance of machine learning classifiers for cancer diagnosis using augmented data.

Main Methods:

• Developed the KAN-AAE method, combining Kolmogorov-Arnold Networks (KAN) and Adversarial Autoencoders (AAE) for synthetic SERS spectra generation.
• Collected SERS spectra from serum samples of cancer patients, other disease patients, and healthy individuals.
• Trained the KAN-AAE model on existing SERS data and generated synthetic spectra, which were combined with real data for classifier training.

Main Results:

• The KAN-AAE method successfully generated high-quality and reliable synthetic SERS spectra.
• Combining synthetic and real data improved classification accuracy by 1% to 3% across various models (logistic regression, decision tree, MLP, 1D-CNN, KAN).
• The KAN classifier achieved the highest accuracy of 95.62% when trained on the augmented dataset.

Conclusions:

• The proposed KAN-AAE method effectively overcomes the challenge of limited data in SERS-based cancer detection.
• Synthetic SERS spectra generation significantly enhances the accuracy and generalization ability of machine learning classifiers.
• This approach holds promise for improving non-invasive cancer diagnosis through advanced spectroscopic and machine learning techniques.