Reading Molecular Signals in Skin with Light: Raman Spectroscopy Combined with AI Achieves 84% Accuracy in Cancer Tissue Classification

The appearance of a suspicious skin lesion can sometimes make it difficult to determine exactly what it is; obtaining tissue and examining it under a microscope can clarify the diagnosis. Reading the tissue’s molecular signals with light first might provide a basis for deciding whether to proceed with a biopsy. A research team at Florida Atlantic University (FAU) in the United States is exploring this approach: combining Raman spectroscopy with machine learning to classify normal skin and two types of skin cancer in excised skin tissue, with a maximum test accuracy of approximately 84%.

The research, announced by FAU on September 30, appears in the Proceedings of SPIE conference proceedings. The team used a portable Raman system equipped with a 785-nanometer diode laser and a handheld probe to obtain nearly 1,000 spectra from more than 50 ex vivo tissue samples, covering normal skin, basal cell carcinoma (BCC), and squamous cell carcinoma (SCC). These spectra were the data analyzed by the models; they do not mean that the study included nearly 1,000 patients.

Raman spectroscopy examines the scattered signals produced when light interacts with tissue molecules, providing clues about chemical composition. News-Medical, citing the FAU research, reported that cancer tissue showed stronger protein-related signals, while normal tissue showed stronger lipid-related signals. The machine learning models’ task was to identify patterns in these subtle differences that could be used for classification.

According to results published by FAU and Medical Xpress, the K-nearest neighbors and support vector machine models achieved the highest overall test accuracy, at approximately 84%; the support vector machine also had a reported sensitivity of 78.7% and specificity of 88.6%. The shallow neural network achieved an accuracy of 80.8%, with an area under the receiver operating characteristic curve (ROC AUC) of 0.910. These metrics reflect different aspects of model performance, and the AUC value cannot be directly interpreted as diagnostic accuracy.

The difficulty of classification also varies. The study found that normal skin was easier to distinguish from cancer tissue, but the molecular signals of basal cell carcinoma and squamous cell carcinoma overlapped more extensively. In other words, spectra can provide clues about abnormal tissue, but the models still face challenges in identifying the specific type of cancer; overall accuracy also cannot fully represent how well each tissue type is identified.

The most significant limitation is that the measurements were taken from excised tissue. The approximately 84% performance cannot yet be considered a measure of diagnostic capability when the probe is placed directly against a patient’s skin, nor has this method been shown to reduce unnecessary biopsies. All three reporting sources relied on research information provided by FAU, making them accounts of the same preliminary findings rather than results independently validated by multiple research teams.

The team next plans to expand the research and optimize the models, including exploring more advanced neural networks. Before the approach can enter clinical practice, researchers must confirm whether these molecular clues are equally reliable in actual lesion assessments and whether they can safely help determine which lesions need further examination. For now, tissue biopsy and pathological examination remain the basis for diagnosis; this study suggests the possibility of adding another tool to support clinical judgment in the future.