Assess whether handheld Raman spectroscopy combined with machine learning can distinguish normal skin, basal cell carcinoma (BCC), and squamous cell carcinoma (SCC).
Approach:
Measurement: Researchers used a mobile system with a 785-nm laser and handheld probe to analyze more than 50 removed tissue samples, generating almost 1,000 Raman spectra.
Analysis: Several machine-learning methods were compared for classifying the spectra into normal skin, BCC, or SCC.
Key Findings:
K-nearest neighbors and support vector machine models achieved approximately 84% overall accuracy; the support vector machine had 78.7% sensitivity and 88.6% specificity.
A shallow neural network achieved 80.8% accuracy and the highest area under the ROC curve (0.910).
Models distinguished normal skin from cancer more successfully than they distinguished BCC from SCC, whose molecular patterns overlapped.
Cancer samples generally showed stronger protein-related signals, while normal skin showed stronger lipid-related signals.
Interpretation:
The study found that Raman spectra contain biochemical information that machine-learning models can use to classify skin tissue. It assessed analytical performance, not clinical diagnostic accuracy.
Limitations:
Testing was performed on removed tissue rather than directly on patients.
The sample set included just over 50 specimens.
Reported performance remained below the greater-than-90% sensitivity, specificity, and accuracy the researchers considered desirable for clinical use.
