Reading Skin’s Molecular Fingerprint with Light: Raman Spectroscopy and AI Show Preliminary Ability to Identify Skin Cancer

Does a suspicious-looking skin lesion necessarily need to be sampled? Benign changes and cancer can sometimes look similar, and physicians still rely on biopsy and microscopic examination for confirmation. A research team at Florida Atlantic University is trying to use light to provide another layer of information: reading chemical signals from tissue and then using machine learning to identify patterns, exploring whether this could eventually help determine which lesions need further examination.

The research, announced by the university on September 30, appears in “Advanced Chemical Microscopy for Life Science and Translational Medicine 2026” in *Proceedings of SPIE*. The team used a mobile Raman system equipped with a 785-nanometer diode laser and a handheld probe to analyze more than 50 clinical tissue samples encompassing basal cell carcinoma, squamous cell carcinoma, and normal skin, obtaining nearly 1,000 spectra. These measurements were taken after the tissue had been removed; the study has not yet demonstrated diagnostic performance directly on patients’ skin.

Raman spectroscopy observes changes in scattering after light interacts with molecules, revealing the tissue’s “molecular fingerprint.” In this set of samples, cancerous tissue generally showed stronger protein-related signals, while normal tissue showed stronger lipid-related signals. The algorithms’ task was to find patterns in these subtle differences that could be used for classification, rather than interpret photographs of lesions.

The study compared several models. K-nearest neighbors and support vector machines achieved the highest overall test accuracy, at about 84%. According to the figures released by the university, the support vector machine had a sensitivity of 78.7% and a specificity of 88.6%; the shallow neural network had an accuracy of 80.8% and an area under the ROC curve of 0.910. These metrics reflect different aspects of classification performance and cannot be directly interpreted as the reliability of ruling out cancer in an outpatient setting.

The difficulty of identification also varied. Reports on the same study by ICT&health and LabMedica both noted that normal skin was easier to distinguish from cancerous tissue, while the molecular signals of the two cancer types overlapped more. The results show that the spectra do contain identifiable tissue information, while also reminding researchers that detecting abnormalities and accurately determining the type of cancer are two different challenges.

Nearly 1,000 spectra do not mean nearly 1,000 participants; the study still included only more than 50 tissue samples, all measured outside the body. The current results therefore cannot yet answer how many cancers this method would miss in an actual patient population, or how many biopsies of benign lesions it could prevent. The three tissue categories included in the study also do not represent all skin lesions that may be encountered in clinical practice.

Senior researcher Andrew Terentis described the results as preliminary findings, and the team hopes to expand the research and optimize the models next. To become an aid for assessment before biopsy, the method will still require measurements taken on patients to be compared with pathology results to establish its reliability and practical benefits. Until then, pathological examination following biopsy remains the diagnostic standard; reducing unnecessary biopsies is a goal this technology hopes to achieve, but has not yet demonstrated.