To evaluate the effectiveness of a handheld Raman spectroscopy system combined with machine learning in distinguishing normal skin, basal cell carcinoma (BCC), and squamous cell carcinoma (SCC).
Approach:
Methodology: Researchers used a mobile Raman spectroscopy system with a 785-nm laser and handheld probe to analyze over 50 removed tissue samples, generating nearly 1,000 Raman spectra.
Machine Learning Models: Various machine-learning methods were compared, including k-nearest neighbors, support vector machines, shallow neural networks, partial least squares discriminant analysis, and principal component analysis with quadratic discriminant analysis.
Key Findings:
K-nearest neighbors and support vector machine models achieved approximately 84% accuracy.
The support vector machine had a sensitivity of 78.7% and specificity of 88.6%.
A shallow neural network achieved 80.8% accuracy with the highest area under the receiver operating characteristic curve at 0.910.
Partial least squares discriminant analysis showed 95.5% specificity but only 66.5% sensitivity.
Normal skin and cancerous tissues were more easily distinguished than BCC from SCC due to overlapping molecular patterns.
Interpretation:
Raman spectra contain biochemical information useful for classifying skin tissue, focusing on analytical performance rather than clinical diagnostic accuracy.
Limitations:
Testing was conducted on removed tissue samples rather than directly on patients.
The sample size was limited to just over 50 specimens.
Reported performance was below the desired clinical thresholds of over 90% sensitivity, specificity, and accuracy.
Conclusion:
The study indicates potential for noninvasive tools in skin cancer diagnosis.
