Improving skin cancer detection by Raman spectroscopy using convolutional neural networks and data augmentation

Area of Science:

• Dermatology
• Medical Spectroscopy
• Artificial Intelligence in Medicine

Background:

• Raman spectroscopy has shown promise for skin cancer detection.
• Previous studies established its good sensitivity and specificity.
• This study investigates enhancing detection via deep neural networks.

Purpose of the Study:

• To determine if combining deep neural networks and Raman spectroscopy improves skin cancer detection.
• To evaluate the impact of data augmentation on diagnostic performance.
• To compare deep learning models against conventional machine learning approaches.

Main Methods:

• Analyzed Raman spectra from 731 skin lesions (340 cancerous/precancerous, 391 benign).
• Developed one-dimensional convolutional neural networks (1D-CNN) for spectral classification.
• Implemented data augmentation (noise, shift, GANs) and compared with PLS-DA, PC-LDA, SVM, LR.

Main Results:

• Data augmentation improved ROC AUC for all models by 2-4%.
• 1D-CNN achieved a ROC AUC of 0.909±0.021 after augmentation, slightly outperforming conventional methods.
• Augmentation improved model performance on noisy or shifted spectra.

Conclusions:

• Data augmentation is crucial for enhancing Raman spectroscopy-based skin cancer detection.
• Deep learning, specifically 1D-CNN, offers a slight advantage over conventional machine learning.
• The combined approach shows potential for improved clinical diagnostic accuracy.