Colorectal cancer (CRC) is one of the most common and deadly forms of cancer worldwide. A key reason for its high death toll is that it often goes undetected until it reaches an advanced stage. Although modern colonoscopy has greatly improved the early detection of precancerous lesions, determining which ones are likely to become cancerous remains a challenge—especially during the procedure itself. A new study published in Biophotonics Discovery may offer a promising solution.
A team led by researchers at the Champalimaud Foundation in Portugal tested an advanced imaging method that uses the natural glow (or autofluorescence) of tissues when exposed to specific wavelengths of light. Importantly, this approach doesn't require any dyes or contrast agents. By analyzing how long this autofluorescence lasts—called "fluorescence lifetime"—the researchers could detect subtle differences in the tissue's biochemistry. Combined with machine learning, this label-free method could distinguish benign from malignant lesions in real time.
The researchers collected fresh tissue samples from 117 patients undergoing colorectal surgery. They scanned each specimen using a fiber-optic probe and a dual-laser autofluorescence lifetime system, which illuminated the tissue at two different wavelengths (375 nm and 445 nm) to excite various molecules like collagen and cellular coenzymes. This setup captured data from multiple optical channels, tracking how each type of tissue responded to the light over time. These measurements were then matched to pathology diagnoses to serve as training data for an AI-based classification model.
Using an ensemble learning technique known as Adaptive Boosting (AdaBoost), the model was trained to recognize patterns in the spectroscopic data that correspond to benign or cancerous tissue. On the training set, the model achieved 87 percent accuracy, with a sensitivity (true positive rate) of 83 percent and a specificity (true negative rate) of 90 percent. When tested on a separate set of new samples, the model performed almost as well, with 85 percent accuracy, 85 percent sensitivity, and 85 percent specificity.
In practical terms, this means the system was able to identify cancerous tissues with high reliability based only on their autofluorescence signatures. Even when applied to individual measurement points—rather than larger tissue regions—the model generated useful probability maps that highlighted tumor areas. These findings suggest that the technology could one day guide doctors during colonoscopy or surgery, helping them decide which lesions to remove and where to focus their attention.
The study also investigated whether the complex imaging system could be simplified without sacrificing accuracy. The results showed that using a reduced set of optical channels—focusing on the most informative biochemical signals—still produced strong performance, potentially lowering the cost and complexity of future clinical devices.
Although more work is needed to improve accuracy, especially for early or borderline lesions, and to expand testing to larger and more diverse patient groups, this research marks an important step toward integrating real-time, AI-guided optical diagnostics into routine colorectal cancer care. It could eventually help reduce unnecessary biopsies, shorten procedures, and ensure that dangerous lesions aren't missed.
For details, see the original Gold Open Access article by J. L. Lagarto et al., " Identification of colorectal malignancies enabled by phasor-based autofluorescence lifetime macroimaging and ensemble learning ," Biophoton. Discovery 2(3) 032705 (2025), doi 10.1117/1.BIOS.2.3.032705
