A handheld Raman spectroscopy system combined with machine learning distinguished normal skin, basal cell carcinoma (BCC), and squamous cell carcinoma (SCC) with up to 84 percent accuracy in a preliminary study published in the Proceedings of SPIE.
BCC and SCC can resemble benign or precancerous skin lesions. Biopsy followed by microscopic examination remains the standard method for diagnosing these cancers, but researchers are investigating noninvasive tools that could help determine which lesions require further assessment.
Raman spectroscopy uses laser light to examine the chemical composition of tissue. When the light interacts with molecules, it produces a spectrum that acts as a molecular fingerprint. Machine-learning models can then search these spectra for patterns associated with cancer.
The researchers used a mobile system equipped with a 785-nm laser and handheld probe. They examined more than 50 removed tissue samples of normal skin, BCC, and SCC, producing almost 1,000 Raman spectra. Several machine-learning methods were compared to determine how accurately they classified the three tissue types.
K-nearest neighbors and support vector machine models achieved the highest overall accuracy, at approximately 84 percent. The support vector machine had a sensitivity of 78.7 percent and specificity of 88.6 percent.
A shallow neural network achieved 80.8 percent accuracy and the highest area under the receiver operating characteristic curve, at 0.910. This measure indicates how well a model distinguishes between diagnostic categories across different classification thresholds.
Other methods showed different strengths and weaknesses. Partial least squares discriminant analysis achieved 95.5 percent specificity but 66.5 percent sensitivity, indicating that it produced relatively few false-positive results but missed more cancer-associated spectra. Principal component analysis with quadratic discriminant analysis had 77.8 percent sensitivity and 77.7 percent specificity.
The models separated normal skin from cancerous tissue more successfully than they distinguished BCC from SCC. The two cancers had overlapping molecular patterns, which contributed to classification errors. Cancer samples generally showed stronger protein-related signals, while normal skin produced stronger lipid-related signals.
The findings show that Raman spectra contain biochemical information that machine-learning models can use to classify skin tissue. However, the study assessed analytical performance rather than clinical diagnostic accuracy.
Testing was performed on removed tissue, not directly on patients, and the sample set included just over 50 specimens. The reported performance also remained below the sensitivity, specificity, and accuracy of more than 90 percent that the researchers considered desirable for clinical use.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
