Diffuse Reflectance Spectroscopy Reached Feasibility for Colorectal Tumor Border Detection
Diffuse reflectance spectroscopy is moving closer to the operating room as a potential tool for colorectal cancer surgery, where the margin between cancerous and healthy tissue can determine both oncologic control and how much function a surgeon can preserve. In fresh colorectal specimens, broadband optical measurements spanning 400 to 1600 nm separated tumor from non-tumor tissue with a level of accuracy that makes the approach hard to dismiss as a laboratory curiosity.
The central problem is familiar to colorectal surgeons: distal resection margins must be wide enough to reduce the chance of leaving residual cancer, yet not so extensive that they compromise sphincter function. Existing intraoperative methods for assessing that border remain limited or cumbersome, leaving no truly rapid tissue-level readout during the procedure. Optical spectroscopy addresses that gap by interrogating tissue composition in real time, using how tissue absorbs and scatters light as a surrogate for microscopic and biochemical differences between malignant and benign regions.
A Large Spectral Dataset and a Binary Tissue Classification Task
Investigators collected continuous measurements from freshly excised colorectal specimens and built a binary classification framework to distinguish tumor from non-tumor locations. The dataset was substantial for this type of translational work: 39,741 spectra drawn from 86 patients, including 22,451 non-tumor spectra and 17,290 tumor spectra. That scale matters because optical methods can appear impressive on small datasets yet falter when they encounter the heterogeneity of real patient tissue.
Several feature extraction strategies were tested alongside multiple classifiers, with hyperparameter tuning applied to optimize performance. Among the approaches evaluated, the first derivative of the spectra combined with a radial basis function support vector machine performed best. Across patients, the mean Matthews Correlation Coefficient reached 0.76 ± 0.27, while sensitivity was 0.93 ± 0.13 and specificity was 0.88 ± 0.17.
Those numbers point to a system that is especially good at finding tumor tissue while also maintaining a relatively strong ability to exclude healthy tissue. In a surgical setting, high sensitivity is essential because a missed tumor border can translate into a positive margin, but specificity is equally important because false positives could drive unnecessary tissue removal and erode functional preservation.
Why Optical Feedback Could Change Margin Assessment
The biological logic behind the technique is straightforward even if the engineering is demanding. Malignant tissue differs from normal colorectal tissue in cellular density, vascularity, architecture, and the way it interacts with light. Diffuse reflectance spectroscopy captures those differences through patterns in reflected light, and machine learning can translate the spectral signatures into a classification decision fast enough to be clinically useful if the hardware is integrated into the workflow.
A broadband system extending from 400 to 1600 nm offers a wider window into tissue behavior than a narrow-band approach. The added spectral range can improve discrimination by sampling multiple absorption and scattering features rather than relying on a single optical cue. In practical terms, that gives the algorithm more information to separate tumor from non-tumor tissue, which is exactly what surgeons need when they are trying to define a resection line in real time.
Still, this is feasibility work, not a finished clinical instrument. The measurements were made on freshly excised specimens rather than during live surgery, so the operating room will introduce additional complexity: motion, blood, smoke, variable contact pressure, and the challenge of positioning a probe consistently at the margin. Each of those factors can distort optical readings and will need to be addressed before the method can guide intraoperative decisions with confidence.
The First Derivative and the RBF Model Carried the Best Signal
The strongest-performing configuration relied on the first derivative of the spectral data paired with a radial basis function support vector machine. Derivative processing can sharpen subtle shape changes in spectra and reduce the influence of broad background shifts, which may explain why it outperformed other feature representations in this cohort. In tissue optics, preprocessing is not a technical footnote; it can determine whether a biologically meaningful signal emerges from a noisy measurement stream.
The patient-level spread in performance also deserves attention. The reported MCC of 0.76 ± 0.27 suggests respectable average discrimination, but the variability signals that some cases were likely much easier than others. In colorectal surgery, that heterogeneity may reflect differences in tumor biology, local tissue composition, specimen handling, or spectral sampling conditions. A clinical device will need to perform well not just on average, but across the full range of patient and tumor characteristics encountered in routine care.
The combination of sensitivity at 0.93 ± 0.13 and specificity at 0.88 ± 0.17 is encouraging because it suggests the tool may one day support margin assessment from both directions: catching tumor when it is present and limiting unnecessary excision when it is absent. That balance is exactly what a margin-assessment technology must achieve to be useful in sphincter-preserving surgery.
Distal Margin Assessment Still Needs Prospective Validation
The immediate next step is prospective validation of distal margin assessment in actual surgical workflows. Specimen-based feasibility is an essential foundation, but it cannot answer the most important operational question: whether a probe can reliably inform the surgeon before the rectum or colon is transected. The future utility of the platform depends on whether the signal remains stable under intraoperative conditions and whether it can be applied at clinically relevant depths.
Probe design will matter as much as the classifier. The study points toward the need for deeper penetration to improve intra-operative applicability, which speaks to a basic limitation of light-tissue interaction in colorectal surgery. Superficial sampling may be enough to detect obvious surface abnormalities, yet a margin assessment tool must remain informative when tumor extends beyond the immediate surface or when tissue layers vary in composition.
There is also a broader translational issue. Machine learning performance in spectroscopy is only as good as the training data, the calibration strategy, and the consistency of the measurement setup. A model that works in one center with one device configuration must still prove itself when moved to a different surgical team, a different patient population, and a different version of the hardware. Until then, the technology remains a promising adjunct rather than a replacement for pathology or surgical judgment.
What this study establishes is feasibility: a broadband optical system, paired with machine learning, can distinguish tumor from non-tumor colorectal tissue with strong classification performance in ex vivo specimens. The remaining work is more exacting. Prospective surgical validation, deeper tissue sampling, and device refinement will determine whether diffuse reflectance spectroscopy can move from an experimental signal to a dependable intraoperative margin tool.