Hyperspectral imaging in medicine: what the 2025 evidence says about surgery, wounds and cancer detection

A diagnostic technology that started life on satellites is moving, slowly but steadily, into operating theatres. Hyperspectral imaging (HSI) lets a camera capture not just the colour of tissue, but its chemical signature, frame by frame. Two major reviews published in 2025, one in August covering 2,163 surgical patients and one in December covering the wider cancer-detection field, suggest the technology is now close enough to clinical use that hospitals, including in the UK, should be paying attention.

This is a non-technical round-up for readers who want to understand what hyperspectral imaging actually is, where it is being trialled today, and what is still holding it back from routine clinical adoption.

Hyperspectral imaging splits light into dozens or hundreds of narrow bands rather than just red, green and blue. Each pixel becomes an optical fingerprint of the tissue beneath it. Photo: Pawel Czerwinski / Unsplash.

What hyperspectral imaging actually is

A normal phone camera sees in three colour bands: red, green and blue. A hyperspectral camera sees in dozens or hundreds of much narrower bands, stretching from visible light (380 to 700 nm) into the near-infrared (800 to 2,500 nm) and sometimes the mid-infrared. Every pixel in a hyperspectral image is not one colour value but a small spectrum, a kind of optical fingerprint.

Because different tissues, blood-oxygen levels, and tumour types absorb and reflect light differently across those bands, software can learn to tell them apart. The technique is non-invasive, requires no contrast dye, uses no ionising radiation, and works in real time. That combination is why surgeons and dermatologists keep coming back to it.

The technology was developed for remote sensing and aerial surveillance in the 1980s. The first serious medical trials began in the early 2000s. The current generation of clinical systems pairs the camera with a deep-learning classifier that turns the spectral fingerprints into a colour-coded map of "this looks like tumour, this looks like healthy tissue" overlaid on the surgeon's view.

What the 2025 evidence actually says

The most comprehensive recent picture comes from a systematic review published on 6 August 2025 in PMC (Liotta and colleagues), which pooled 2,163 patients across 82 separate studies, including 24 paediatric cases from three neurosurgery trials. The review covers six of the cancer settings where HSI has been tested most aggressively:

• Neurosurgery (30 studies): glioblastoma, low-grade glioma, meningioma, pituitary adenoma
• Head and neck (22 studies): squamous cell carcinomas of the pharynx, larynx and oral cavity
• Breast (12 studies): invasive ductal carcinoma and ductal carcinoma in situ
• Colorectal (8 studies) and gastroesophageal (7 studies)
• Other: pancreatic, hepatocellular, thyroid, kidney and ovarian

The headline performance numbers are good. For tumour segmentation against healthy tissue, a support-vector-machine classifier in a preclinical mouse model reached 93.7% sensitivity and 91.3% specificity. Breast-tissue classification accuracy in ex vivo specimens sat at 93 to 98% for invasive carcinoma and ductal carcinoma in situ. For glioma biopsy diagnosis under fluorescence, the hyperspectral system achieved an area under the curve of 0.845, beating the standard surgical microscope's 0.710.

The review's overall verdict is cautiously optimistic. In the authors' own words: "HSI remains in its early technological stages, requiring high-quality evidence and multidisciplinary collaboration to support clinical adoption."

Where it is being used in practice

Real-time intraoperative HSI gives the surgeon a colour-coded map of tumour versus healthy tissue, generated in around a minute. Photo: Piron Guillaume / Unsplash.

Brain surgery: the HELiCoiD project

The European HELiCoiD ("HypErspectraL Imaging Cancer Detection") project, which ran from 2016 to 2021 and continues to generate follow-on research, is the most cited clinical application. The project released the first open in vivo brain hyperspectral database, with over 300,000 labelled spectral signatures drawn from 36 images of 22 patients. An extended version added 62 images from 34 patients spanning brain tumour grades I to IV plus secondary lesions from breast, lung and kidney cancers.

In real intraoperative use, the HELiCoiD-derived system has shown that it can discriminate between healthy brain tissue and tumour tissue in roughly one minute during the procedure. That matters because the surgeon's central question, "have I taken enough?", is the one neurosurgical complication patients pay for in lost cognitive function.

Diabetic foot ulcers

This is the application closest to routine clinical use. A study reported in PubMed (PMID 36633904) used oxyhaemoglobin measurements from a hyperspectral camera to predict whether a diabetic foot ulcer would heal. At the first visit, the system achieved 85% sensitivity and 70% specificity for healing prediction. By the second visit, specificity rose to 85%. Predicting non-healing early matters because it triggers earlier escalation to vascular intervention, which is the difference between a recovered foot and an amputation.

Endoscopic cancer detection

A December 2025 Frontiers in Immunology editorial (Wu, Wang and Mukundan) collected several recent advances in multispectral and hyperspectral imaging for cancer detection. The most striking number from that round-up: AI-enhanced hyperspectral endoscopy improved oesophageal cancer detection accuracy by 8% versus conventional white-light endoscopy. The editorial also covers breast lesion characterisation, adrenal melanoma diagnosis, lung cancer differentiation, prostate cancer early detection, and spinal compression fractures.

Wound assessment and burn triage

Major US institutions including Mayo Clinic, Stanford Medicine and Cleveland Clinic have run feasibility studies on hyperspectral imaging for burn-depth assessment and wound diagnostics. These are the studies driving the current commercial market: a hyperspectral image of a burn can distinguish superficial from deep partial-thickness injury within seconds, where clinical judgement alone is correct only around 60 to 70% of the time.

What is still holding HSI back

The August 2025 review names six specific barriers honestly. UK clinicians and procurement teams evaluating this technology should weigh all of them.

1. Image acquisition is slow. Capturing a single hyperspectral image still takes 30 to 60 seconds, which disrupts the surgical workflow and prevents large in vivo training datasets from being built quickly.
2. Blood contamination causes false positives. Haemoglobin's absorption spectrum is strong enough to confuse the classifier at tumour margins. Robust handling strategies are still being developed.
3. There is no agreed standard. Different hardware, different processing pipelines, different study designs, all of which means published accuracy figures cannot be compared directly.
4. Real-time processing is not quite real-time yet. Calibration and fluorescence processing both add lag.
5. Paediatric data is sparse. Algorithms trained on adult tissue may not generalise.
6. The hardware is complex. Different sensor types (CCD, CMOS, InGaAs) and multiple light sources are needed to cover the full useful spectrum, which keeps system costs high and integration awkward.

The authors of the review summarise the situation cleanly: "a deeper understanding and improved characterisation of biological tissue hyperspectral properties are essential to better inform and orient future hardware and software designs."

Diabetic foot ulcer assessment is the application closest to routine clinical use. An early "will this heal?" answer can be the difference between recovery and amputation. Photo: National Cancer Institute / Unsplash.

Why this matters for UK readers

NHS surgical oncology and diabetic-foot services are exactly the high-stakes, time-pressured settings where a one-minute "is this cancer" answer or a thirty-second "will this ulcer heal" answer would change clinical pathways. The technology is not approved for routine use yet, and no single device has cleared full UK regulatory adoption. But the 2025 evidence base now has enough patients, enough cancer types, and enough peer review behind it that clinical-research teams should be planning trials, not waiting for the next review.

For non-clinical readers, the broader story is simpler. Hyperspectral imaging is one of several "AI-plus-physics" diagnostic technologies (alongside near-infrared spectroscopy, photoacoustic imaging, and Raman spectroscopy) that are moving from physics-laboratory curiosities into the operating theatre. The 2026 to 2028 window is likely to be when several of them either prove themselves or quietly stall. Hyperspectral imaging looks, on current evidence, like a strong candidate to be one of the ones that proves itself.

Sources

• Hyperspectral imaging for tumor resection guidance in surgery: a systematic review of preclinical and clinical studies, Liotta et al., PMC, 6 August 2025.
• Editorial: Recent trends and advancements in multispectral and hyperspectral imaging for cancer detection, Wu, Wang and Mukundan, Frontiers in Immunology, 8 December 2025.
• Advancing hyperspectral imaging and machine learning tools toward clinical adoption in tissue diagnostics: a comprehensive review, APL Bioengineering, December 2024.
• HELICoiD project: a new use of hyperspectral imaging for brain cancer detection in real-time during neurosurgical operations, SPIE proceedings, project active 2016 to 2021 with continuing follow-on research.
• Use of hyperspectral imaging to predict healing of diabetic foot ulceration, PubMed (PMID 36633904).
• Hyperspectral imaging in neurosurgery: a review of systems, computational methods, and clinical applications, PMC, 2024.