During pancreatic cancer surgery, removing all tumor tissue while preserving healthy pancreas and surrounding structures is critically important. Leaving any cancer cells at the margin of the resection dramatically worsens outcomes. Currently, surgeons rely on intraoperative frozen-section analysis (IFSA) — rapidly freezing and slicing tissue for pathological examination — but this technique has sensitivity as low as 33% and takes valuable operating time.
Other intraoperative tools like ultrasound and near-infrared imaging have significant limitations in sensitivity, field of view, or the need for injectable fluorescent dyes. A rapid, label-free technology that could be pointed at tissue and instantly classify it as healthy, inflamed, or cancerous would transform pancreatic surgery.
Raman spectroscopy works by shining a laser on a tissue sample and measuring the light that scatters back. Different molecules vibrate at different frequencies, so the scattered light pattern — called a Raman spectrum — acts as a unique chemical fingerprint for that tissue. Cancer cells have different protein, lipid, and nucleic acid compositions than healthy or inflamed cells, producing measurably distinct spectra.
This pilot study collected Raman spectra from 15 patients' pancreatic tissue samples, classifying each as healthy pancreas (N), pancreatitis (P), or ductal adenocarcinoma (DA). The researchers developed a novel spectral preprocessing method called SPectral SELection (SPSEL) to remove noise and irrelevant spectral information, then applied principal component analysis (PCA) to reduce the data dimensions before final classification.
Two machine learning classifiers were tested on the PCA-reduced Raman data: Gaussian Naive Bayes (GNB) and Random Forest Classifier (RFC). These algorithms were trained to recognize which spectral patterns correspond to each of the three tissue classes. The SPSEL preprocessing step was critical — without it, background noise and overlapping spectral regions would confuse the classifiers.
The combination of PCA dimensionality reduction followed by Random Forest classification on SPSEL-preprocessed spectra achieved the highest accuracy, reaching up to approximately 96%. This means the algorithm correctly classified nearly all tissue samples as healthy, pancreatitis, or cancer.
Achieving 96% accuracy across three tissue classes — and specifically being able to separate cancer from pancreatitis — is clinically significant. Pancreatitis (pancreatic inflammation) produces tissue changes that look similar to cancer under many imaging modalities, making it one of the most common diagnostic pitfalls in pancreatic disease. Raman spectroscopy, guided by machine learning, navigates this challenge effectively.
The label-free, non-destructive nature of Raman spectroscopy is particularly valuable for surgical applications — the same tissue sample can be measured and then handed to pathology for gold-standard confirmation. No special preparation, no dyes, no tissue destruction required.
This pilot study, while small at 15 cases, provides strong proof-of-concept that Raman spectroscopy combined with machine learning can achieve near-expert accuracy in pancreatic tissue classification. The technology is fast enough for intraoperative use — measurements take seconds — and requires no consumables beyond the laser probe itself.
Scaling to larger multi-center studies and developing miniaturized, handheld Raman probes are the next steps toward clinical deployment. If confirmed in larger trials, this could become the first real-time molecular margin assessment tool in pancreatic cancer surgery, directly improving the completeness of tumor removal.