AI and light could help reduce unnecessary skin biopsies

Researchers at Florida Atlantic University (FAU) are investigating whether a combination of light-based tissue analysis and machine learning could help distinguish skin cancer from healthy tissue without immediately taking a biopsy. In an initial study, the best-performing algorithms achieved an accuracy of around 84 percent.

The approach combines Raman spectroscopy with artificial intelligence. Raman spectroscopy uses the way light interacts with molecules to provide information about the chemical composition of tissue. Rather than removing tissue for laboratory analysis, the technique can effectively create a molecular fingerprint. The research is still at an early stage and is not intended to replace conventional pathology. However, it could eventually provide clinicians with an additional tool for deciding which suspicious skin lesions require a biopsy.

Reading the molecular fingerprint

Distinguishing malignant skin lesions from benign or precancerous abnormalities can be difficult because they can look similar. A biopsy followed by microscopic examination remains the gold standard for diagnosing skin cancer. The drawback is that biopsies are invasive and some ultimately reveal that a suspicious lesion is benign. The FAU researchers therefore explored whether molecular differences could provide additional information before tissue is removed. They used a mobile Raman spectroscopy system with a handheld probe and analysed more than 50 ex vivo clinical samples, including basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and normal skin.

Almost 1,000 Raman spectra were generated and analysed using several machine-learning methods. K-nearest neighbours and support vector machine classifiers achieved the highest overall test accuracy at approximately 84 percent. The support vector machine reached 78.7 percent sensitivity and 88.6 percent specificity. A shallow neural network achieved 80.8 percent accuracy and the highest ROC AUC, at 0.910. Normal tissue could be distinguished relatively well from cancerous tissue, while BCC and SCC showed more overlap. Cancer samples generally produced stronger protein-related signals, whereas normal tissue showed stronger lipid-related signals.

AI at different stages of diagnosis

The study adds another potential application to the growing role of AI in skin cancer diagnostics. Earlier research (2025) showed how AI can also support specialists after tissue has already been removed. In that international study, led by researchers at Karolinska Institutet in collaboration with Yale University, AI was used to support the assessment of tumour-infiltrating lymphocytes, or TILs, in melanoma tissue. These immune cells provide information about the immune response around a tumour and can contribute to assessing prognosis and treatment. The retrospective study involved 98 pathologists and other researchers assessing 60 tissue sections from patients with malignant melanoma.

AI-supported quantification made assessments more consistent and improved prognostic accuracy compared with conventional assessment. The researchers argued that AI could make an otherwise partly subjective pathological assessment more objective, although further validation was required before clinical implementation. The two approaches target different parts of the diagnostic pathway. AI-assisted Raman spectroscopy could potentially help determine whether a lesion warrants further investigation before biopsy, while AI-supported digital pathology could make the subsequent microscopic assessment more consistent.

From promising results to clinical use

For the Raman approach, considerable work remains before that scenario becomes clinical reality. The current study involved a relatively small number of samples analysed outside the body. The researchers plan larger studies and further optimisation of the algorithms, including the use of deep neural networks. According to senior author Andrew Terentis, the objective is to develop technology that is not only accurate but also portable and practical enough for clinical use.

The significance of the study therefore lies less in the current accuracy figure than in the proposed diagnostic workflow. If larger clinical studies confirm the findings, molecular information from light could become an additional layer between visual examination and biopsy. Combined with AI, that could help clinicians identify which lesions require tissue analysis while potentially avoiding biopsies when the evidence indicates that they are unlikely to be necessary.