Reading Skin Cancer’s Molecular Fingerprint

Machine-learning models analyzed Raman spectra to distinguish cancerous tissue from normal skin

October 6, 2026
2 min

The Analytical Scientist

Overview

A handheld Raman spectroscopy system combined with machine learning demonstrated up to 84% accuracy in distinguishing normal skin, basal cell carcinoma (BCC), and squamous cell carcinoma (SCC).

Background

Skin cancers such as BCC and SCC can mimic benign or precancerous lesions, making accurate diagnosis challenging. Traditional biopsy and microscopic examination remain the gold standard for diagnosis, but there is a growing interest in noninvasive methods that could streamline the assessment process. The development of technologies like Raman spectroscopy may provide valuable biochemical insights for more accurate tissue classification.

Data Highlights

Model	Accuracy	Sensitivity	Specificity
K-nearest neighbors	84%	78.7%	88.6%
Support vector machine	84%	78.7%	88.6%
Shallow neural network	80.8%	N/A	N/A
Partial least squares discriminant analysis	N/A	66.5%	95.5%
Principal component analysis with quadratic discriminant analysis	N/A	77.8%	77.7%

Key Findings

The handheld Raman spectroscopy system achieved up to 84% accuracy in classifying skin tissue types.
K-nearest neighbors and support vector machine models provided the highest accuracy, with a sensitivity of 78.7% and specificity of 88.6%.
Partial least squares discriminant analysis had high specificity (95.5%) but lower sensitivity (66.5%).
Normal skin and cancerous tissues showed distinct biochemical signals, aiding in classification.
Classification errors were more frequent between BCC and SCC due to overlapping molecular patterns.

Clinical Implications

Further validation in clinical settings is necessary to establish the effectiveness of Raman spectroscopy compared to traditional biopsy methods.

Conclusion

This study indicates that Raman spectroscopy can classify skin tissue types noninvasively, though its clinical diagnostic accuracy requires further investigation.