Cancer tissue has a different molecular signature than healthy tissue.
In the operating theater, the line between what must be removed and what must be preserved has long been drawn by imperfect human sight. Researchers at Imperial College London have now demonstrated that laser light, reading the molecular signatures of tissue and interpreted by machine learning, can locate that line with over 97% accuracy—distinguishing not only cancer from healthy breast tissue, but one form of cancer from another. The work points toward a future in which surgeons receive real-time guidance during breast-conserving procedures, potentially sparing patients the burden of repeat operations and the quiet losses that come with them.
- Every year, a significant share of breast-conserving surgeries end with a second operation because cancerous cells were left behind at the margins—a problem this technology is designed to eliminate.
- The system achieved over 97% accuracy in the binary task of separating cancer from healthy tissue, and between 83 and 96% sensitivity when classifying four distinct tissue subtypes, including pre-invasive disease.
- The technique is label-free, requiring no chemical staining or preparation, which is precisely what makes real-time intraoperative use conceivable rather than theoretical.
- The research remains in the laboratory phase, and the path from controlled tissue samples to the noise and urgency of a live operating room will require rigorous clinical validation before surgeons can rely on it.
Surgeons performing breast-conserving operations face a fundamental uncertainty: the boundary between tumor and healthy tissue is not always visible to the eye. Remove too little and cancer remains; remove too much and the patient loses tissue unnecessarily, sometimes requiring a second surgery. A study from Imperial College London and collaborating institutions now suggests that Raman spectroscopy—a laser-based technique that reads the molecular composition of tissue—combined with machine learning, can resolve this uncertainty with remarkable precision.
The research team analyzed 80 tissue samples from 71 patients, scanning each with a confocal Raman microscope and feeding the spectral data into algorithms trained to detect patterns beyond human perception. In distinguishing cancer from normal tissue, the system achieved 97.84% sensitivity and 97.18% specificity. When pushed to classify four tissue categories—healthy tissue, invasive ductal carcinoma, invasive lobular carcinoma, and ductal carcinoma in situ—it reached sensitivity between 83 and 96% depending on type, with specificity as high as 99%.
The clinical significance lies not only in the accuracy but in what the system can differentiate: invasive cancers that have spread into surrounding tissue versus pre-invasive disease still confined to the ducts. These distinctions carry real consequences for treatment decisions. Crucially, the technique requires no staining or chemical preparation, making it compatible with the time constraints of surgery.
Conducted under ethical oversight with informed patient consent and funded by the National Institute for Health Research, the work is still in its laboratory phase. But if these results translate to the operating room, surgeons could one day scan tumor margins in real time, receiving immediate confirmation that an excision is complete—reducing repeat procedures, time under anesthesia, and the physical and emotional toll of a second operation.
Surgeons removing breast tumors face a persistent problem: knowing exactly where the cancer ends and healthy tissue begins. Cut too little, and cancer cells remain. Cut too much, and the patient loses unnecessary breast tissue and may need a second operation. A new study suggests that Raman spectroscopy—a technique that reads the molecular fingerprint of tissue using laser light—combined with artificial intelligence, can solve this problem with striking accuracy.
Researchers at Imperial College London and collaborating institutions tested whether Raman spectroscopy could distinguish not just between cancerous and normal breast tissue, but also between different types of breast cancer. They collected 80 tissue samples from 71 patients: 46 samples of healthy breast tissue and 34 samples of cancer. Using a confocal Raman microscope, they scanned each sample and fed the resulting spectral data into machine learning algorithms trained to recognize patterns invisible to the human eye.
The results were compelling. When the algorithm was asked to simply sort tissue into two categories—cancer or not cancer—it succeeded 97.84% of the time at identifying cancer and 97.18% of the time at correctly identifying healthy tissue. But the researchers pushed further. They trained the system to distinguish among four categories: normal tissue, invasive ductal carcinoma (the most common form of breast cancer), invasive lobular carcinoma (a less common invasive type), and ductal carcinoma in situ (a pre-invasive form that has not yet penetrated the duct walls). For this more demanding task, the algorithm achieved sensitivity ranging from 83 to 96% depending on the cancer type, with specificity between 93 and 99%.
What makes this work clinically relevant is that it captures a distinction surgeons care deeply about: the difference between invasive disease that has spread into surrounding tissue and pre-invasive disease confined to the ducts. These require different treatment decisions. The technique is also label-free, meaning it requires no staining or chemical preparation of the tissue—a crucial advantage if the goal is to use it during surgery, when speed matters.
The study was conducted under strict ethical oversight, with tissue samples collected through the Imperial College Healthcare Tissue Bank and Breast Cancer Now, a registered charity. All participants gave informed consent, and the work followed international guidelines for clinical research. The research received funding from the National Institute for Health Research and the Hong Kong government's innovation initiative.
If these laboratory results hold up in the operating room, the implications are substantial. A surgeon could potentially use Raman spectroscopy to scan the margins of a tumor in real time, getting immediate feedback on whether the excision is complete. This could reduce the number of patients who need a second surgery to remove residual cancer—a common outcome of breast-conserving surgery today. For patients, that means fewer procedures, less time under anesthesia, and better preservation of breast tissue and appearance. The technology is not yet in clinical use, but the accuracy demonstrated here suggests it may soon move from the research lab toward the operating theater.
Citações Notáveis
RS can accurately distinguish normal from cancerous tissue and capture clinically relevant differences among histological subtypes, including invasive and pre-invasive disease.— Study authors
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that the algorithm can tell the difference between invasive and pre-invasive cancer? Aren't they both cancer?
They are, but they're treated very differently. Pre-invasive cancer hasn't broken through the duct wall yet, so it's contained. Invasive cancer has spread into surrounding tissue and is more dangerous. A surgeon needs to know which one they're dealing with to decide how much tissue to remove and what follow-up treatment to recommend.
The accuracy numbers are very high—97% for cancer detection. Is that realistic, or does it only work in a lab?
That's the honest question. These are ex vivo samples—tissue removed from the body and measured on a microscope. The real test will be whether it works during actual surgery, when tissue is still in the patient, blood is present, and time pressure is real. But the fact that it works this well on removed samples is a necessary first step.
How does Raman spectroscopy actually work? What is it reading?
It bounces laser light off tissue and measures how the light scatters back. Different molecules—proteins, lipids, collagen—scatter light in different ways. Cancer tissue has a different molecular composition than healthy tissue, so it produces a different spectral signature. The AI learns to recognize those signatures.
If a surgeon had this tool during surgery, what would they actually do with it?
They'd scan the edge of the tumor they just removed. If the algorithm says "cancer detected," they know they need to cut deeper or wider. If it says "healthy tissue," they can stop. Right now, surgeons rely on visual inspection and feel, which is imperfect. This would give them real-time feedback.
How many patients would benefit from this?
Breast-conserving surgery is the standard treatment for early-stage breast cancer in many countries. Tens of thousands of women have it each year. If this reduces re-excision rates even modestly, that's a significant reduction in unnecessary surgery and its physical and psychological costs.