One Raman Scan Separates Oral Cancer From Healthy Tissue With 87.5% Accuracy
Oral cancer is usually confirmed only after a scalpel does the work — this study shows a laser can flag it first.
Oral cancer is the sixth most common malignancy worldwide, and squamous cell carcinoma accounts for roughly 90% of the 300,000 cases diagnosed each year. About 60% of new cases and 68% of oral-cancer deaths occur in Asia. Despite that burden, oral squamous cell carcinoma (OSCC) is still usually caught late — the overall five-year survival rate sits around 50%, largely because the gold-standard diagnostic, biopsy, is invasive, requires an incision, and isn’t something clinicians reach for at every visit.
The Obstacle — Screening Needs to Be Fast, Painless, and Objective
The tools available for catching oral cancer earlier — vital staining, light-based visual screening, other optical techniques — help, but none of them alone reliably identifies oral potentially malignant disorders (OPMDs) before they progress. Biopsy remains definitive, but its invasiveness limits how often and how broadly it can realistically be used for screening. What’s missing is a fast, non-destructive way to interrogate tissue chemistry directly and flag malignancy objectively, without waiting for a pathology report.
The Innovation — Classifying Tissue by Its Molecular Fingerprint
In a study published in the Journal of Clinical Medicine (Jeng, Sharma, Sharma, Chao, Huang, Chang, Wu & Chow, 2019, DOI: 10.3390/jcm8091313), researchers at Chang Gung University and Chang Gung Memorial Hospital, with a collaborator at the University of Central Florida, tested whether Raman spectroscopy (RS) could distinguish oral tumor tissue from healthy tissue using only its molecular vibrational signature — no stain, no label, no incision beyond the surgical excision already being performed.
The team collected 80 cryopreserved tissue samples (44 tumor, 36 normal) from three oral sub-sites — tongue, buccal mucosa, and gingiva — during surgery at Chang Gung Memorial Hospital between February 2017 and December 2018 (IRB No. 201800420B0). Samples were flash-preserved in liquid nitrogen at −80°C immediately after excision to lock in their native molecular structure before Raman measurement. Each sample was scanned five times at different points, using a Raman system built by ProTrusTech Co., Ltd. (Taiwan) with a 532 nm excitation laser, a 5-second integration time, a 15-second acquisition time, and each displayed spectrum averaged from three scans — yielding 400 individual spectra for point-wise analysis and 80 patient-averaged spectra for patient-wise analysis.
Rather than relying on a single Raman peak, the team fed the full spectral fingerprint (700–2000 cm-1) through principal component analysis (PCA) to reduce dimensionality, then classified samples using two supervised models — linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) — validated by both k-fold and leave-one-out cross-validation (LOOCV).
Key Results — From a Spectral Signature to a Working Classifier
Normal tissue spectra were lipid-dominated, with peaks at 1004, 1156, 1339, 1450, 1523, and 1656 cm-1. Tumor tissue spectra were protein-dominated, with peaks at 754, 1064, 1168, and 1220 cm-1 — consistent with elevated protein, amino acid, and nucleic acid content in malignant tissue.
Patient-Wise PCA-QDA Reached 87.5% Accuracy
The patient-wise PCA-QDA model reached 87.5% accuracy (sensitivity 90.90%, specificity 83.33%) distinguishing tumor from normal tissue across all three sub-sites combined, outperforming patient-wise PCA-LDA (81.25% accuracy, 77.27% sensitivity, 86.11% specificity). Averaging each patient’s five spectra into one — rather than classifying all 400 individual point spectra separately, which reached only 74.5–81.75% accuracy — measurably strengthened the result. Cross-validation confirmed it wasn’t overfit: the PCA-QDA error rate held at 11.25% (LOOCV) and 16.25% (k-fold), closely matching the original 12.5% test-split error.
Sub-Site Accuracy Peaked at 91.3% for Gingiva Tissue
Broken down by sub-site, patient-wise PCA-LDA reached 91.30% accuracy at the gingiva sub-site, while PCA-QDA reached 87.5% accuracy (100% sensitivity) at the tongue sub-site and 87.87% accuracy at the buccal mucosa sub-site. Tongue tissue showed the lowest specificity (72.72%), which the authors attribute to its layered structure creating overlap between healthy sub-layers and tumor tissue.
Why ProTrusTech Is Featuring This Research
The Raman measurements in this study — every one of the 400 spectra behind the classification models — were collected on a Protrustech MRI Raman spectroscopy system in a study co-authored across Taiwan and the U.S. — Chang Gung University/Memorial Hospital with the University of Central Florida.
This isn’t a case of a research group using our instrument to verify one supporting figure; the entire dataset the paper’s diagnostic accuracy claims rest on came from our hardware, in a clinical hospital setting, on real patient tissue.
That’s a meaningfully different bar than a materials-science measurement, and it’s exactly the kind of validation — real diagnostic accuracy, on human tissue, published in a peer-reviewed clinical journal — that matters most to us.
Featured Instruments:
Why This Matters
Beyond oral cancer specifically, this paper is a proof point for a broader idea: a tissue sample’s Raman spectrum carries enough biochemical information — protein, lipid, and nucleic acid content — to support an objective, machine-learned malignancy classification, without a stain, a label, or additional invasive steps beyond the excision itself. That principle isn’t limited to oral tissue or to cryopreserved samples; the authors note their next step is pairing this kind of spectral data with more advanced machine learning to further reduce the error rate.
What’s Next
- Larger sample sizes per sub-site, since the current counts (tongue: 24, buccal: 33, gingiva: 23) are a limiting factor the authors note explicitly.
- Machine learning beyond LDA/QDA, applying meta-learning and neural network approaches to the same spectral dataset to push error rates down further.
- Intra-operative margin assessment, using Raman for real-time evaluation of surgical margins to flag residual tumor tissue during surgery rather than waiting for post-operative pathology.
Evaluating Raman spectroscopy for a clinical diagnostics or tissue-classification application of your own? Get in touch with our team.
Key Takeaways
- A ProTrusTech Raman system collected all 400 spectra behind this study’s oral cancer classification models — no staining or labeling required.
- The patient-wise PCA-QDA model reached 87.5% accuracy (90.90% sensitivity, 83.33% specificity) distinguishing tumor from normal oral tissue across tongue, buccal mucosa, and gingiva samples.
- Normal tissue spectra were lipid-dominated (1004, 1156, 1339, 1450, 1523, 1656 cm-1); tumor tissue spectra were protein-dominated (754, 1064, 1168, 1220 cm-1).
- Averaging each patient’s spectra (patient-wise) outperformed classifying individual point spectra, and the result held up under both k-fold and leave-one-out cross-validation.
- The authors’ next steps are larger sample sizes, neural-network-based classification, and real-time intra-operative margin assessment.
