📄 medicine

Feasibility of Diffuse Reflectance Spectroscopy for Distal Margin Assessment in Colorectal Cancer Surgery

This study demonstrates the feasibility of using broadband diffuse reflectance spectroscopy combined with machine learning to accurately distinguish between tumor and healthy tissue in colorectal cancer specimens, achieving high sensitivity and specificity for potential intra-operative distal margin assessment.

Original authors: Joris T. Hepkema, Isis Boeije, Arend G.J. Aalbers, Brechtje A. Grotenhuis, Marinke Westerterp, Freija Geldof, Behdad Dashtbozorg, Theo J.M. Ruers

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Joris T. Hepkema, Isis Boeije, Arend G.J. Aalbers, Brechtje A. Grotenhuis, Marinke Westerterp, Freija Geldof, Behdad Dashtbozorg, Theo J.M. Ruers

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a surgeon performing a delicate operation to remove a piece of cancerous bowel. Your goal is to cut out the tumor completely but leave as much healthy tissue as possible so the patient can still function normally. The tricky part is knowing exactly where the "bad" tissue ends and the "good" tissue begins. Right now, surgeons often have to guess, mark the spot with ink (which can be messy), or wait hours for a lab test that might still miss the edge.

This paper is like a report on a new, high-tech "magic flashlight" that helps surgeons see the invisible line between cancer and healthy tissue instantly.

The Problem: Finding the Invisible Edge

Think of the bowel as a long tube. The tumor is a stain inside the tube. When you cut the tube, you need to make sure you cut far enough past the stain so no cancer is left behind, but not so far that you cut away too much healthy tube.

Currently, surgeons are flying blind in a way. They can't feel the difference between the cancer and the healthy tissue with their hands (especially in minimally invasive surgery). They sometimes use frozen lab tests, but that takes a long time and isn't perfect.

The Solution: The "Magic Flashlight" (Diffuse Reflectance Spectroscopy)

The researchers tested a device called Diffuse Reflectance Spectroscopy (DRS).

How it works: Imagine shining a very bright, multi-colored flashlight (covering colors from visible light to infrared) onto the tissue. The tissue bounces some of that light back.
The "Fingerprint": Just like every person has a unique fingerprint, cancerous tissue and healthy tissue bounce light back in slightly different patterns. The machine reads this "light fingerprint" instantly.
The Speed: Instead of taking one snapshot every 15 seconds (which is too slow for surgery), this new system takes 5 to 8 "snapshots" every second. It's like scanning a barcode rapidly across the surface of the tissue.

The Experiment: Training the Computer Brain

The team didn't just build the flashlight; they had to teach a computer how to understand the light patterns.

The Data: They took 86 freshly removed bowel specimens from patients. They scanned the cancerous parts and the healthy parts, collecting nearly 40,000 light readings.
The Training: They fed these readings into a computer program (Machine Learning). They tried different ways to process the data, like asking the computer to look at the "slope" of the light changes (the first derivative) rather than just the raw brightness.
The Best Team: They found that the best combination was using the "slope" of the light data paired with a specific type of math model called an RBF Support Vector Machine. Think of this as the computer learning to spot the subtle "wiggles" in the light pattern that only cancer makes.

The Results: A Very Sharp Eye

The computer learned to distinguish cancer from healthy tissue with impressive accuracy:

It correctly identified cancer about 89% of the time.
It correctly identified healthy tissue about 93% of the time.
In a "proof-of-concept" test, they used the system on fresh tissue and placed a stitch to mark where the computer said the tumor ended. When they checked later, the computer's guess was very close to the actual tumor edge.

The Catch (Limitations)

The paper is honest about what this tool can't do yet:

Depth: The flashlight can only "see" about 1 to 2 millimeters deep. If the cancer is hiding deeper inside the wall, the light might miss it.
The Future: The researchers say this is a "proof-of-concept." It works well on tissue that has already been removed. The next step is to prove it works perfectly while the surgeon is actually operating inside a patient's body.

The Bottom Line

This paper shows that using a special light scanner combined with smart computer algorithms is a feasible way to tell the difference between cancer and healthy bowel tissue in real-time. It's like giving the surgeon a superpower to see the invisible edge of a tumor, potentially helping them save more healthy tissue while ensuring the cancer is fully gone. However, the paper stops short of saying this is ready for every hospital tomorrow; it simply proves the idea works in a controlled setting.

Technical Summary: Feasibility of Diffuse Reflectance Spectroscopy for Distal Margin Assessment in Colorectal Cancer Surgery

Problem Statement
In colorectal cancer surgery, particularly for rectal tumors, achieving an adequate distal resection margin (DRM) is critical for oncological safety while preserving sphincter function. Current intra-operative methods for assessing the DRM have significant limitations. Tattooing, used to mark tumor boundaries pre-operatively, suffers from ink spreading and low accuracy (as low as 69%). Frozen section analysis, the current gold standard, is time-consuming, can miss distal tumor spread due to lesion fragmentation after chemoradiation, and relies heavily on surgeon preference. Furthermore, laparoscopic and robot-assisted procedures lack tactile feedback, making the localization of distal tumor boundaries challenging. There is a need for a real-time, non-contact guidance system that can rapidly distinguish tumor tissue from healthy tissue to assist surgeons in determining the optimal transection level.

Methodology
This study employed a prospective, ex vivo design involving 86 patients with colorectal cancer (T3 or higher) at the Netherlands Cancer Institute.

Data Acquisition: A broadband Diffuse Reflectance Spectroscopy (DRS) system (400–1600 nm) utilizing a laser light source and two spectrometers was used. Unlike traditional point-based systems, this setup operated continuously, capturing 5 to 8 spectra per second. Measurements were taken on freshly excised specimens immediately after removal.
- Tumor Measurements: Performed over the gross tumor location.
- Non-Tumor Measurements: Performed on healthy colon/rectal wall and adipose tissue at least 1 cm away from the tumor.
- Volume: A total of 39,741 spectra were collected (22,451 non-tumor, 17,290 tumor).
Ground Truth: Tissue labels were assigned based on histopathologic analysis. Spectra were labeled "colorectal tumor" if tumor tissue was present at the measurement location (regardless of depth) and "non-tumor" otherwise.
Data Preprocessing: Spectra were calibrated against dark and white references, stitched to create a continuous 400–1600 nm range (356 bands), and normalized using Standard Normal Variate (SNV) to correct for intensity variations.
Machine Learning Pipeline:
- Feature Selection: Five strategies were evaluated: (A) Full spectra, (B) First derivative of spectra, (C) LASSO regression, (D) Linear Discriminant Analysis (LDA), and (E) k-means spectral band clustering.
- Classifiers: Five models were tested: Linear SVM, Polynomial SVM, Radial Basis Function (RBF) SVM, Multi-Layer Perceptron (MLP), and Random Forest.
- Training Strategy: Data was split patient-wise (80% training/validation, 20% test) to prevent leakage. Synthetic Minority Oversampling Technique (SMOTE) was used to balance classes. Hyperparameter tuning (grid search for $C$ and $\gamma$ ) was performed over ten iterations with different patient splits to ensure robustness.
Proof-of-Concept Testing: The final optimized model was tested on three fresh specimens. Measurements were performed linearly across the tumor boundary, with real-time feedback provided via an audible signal (pitch change) indicating the predicted tissue type. The predicted boundary was marked with a suture and compared to the macroscopic tumor edge.

Key Results

Model Performance: The optimal configuration utilized the first derivative of the spectra as input features combined with an RBF Support Vector Machine (SVM).
- Matthews Correlation Coefficient (MCC): 0.820 ± 0.015.
- Sensitivity: 0.888 ± 0.015.
- Specificity: 0.928 ± 0.028.
- AUROC: 0.95.
Feature Analysis: Using the first derivative consistently outperformed other feature selection methods (including full spectra and k-means) across most classifiers, likely due to noise reduction and the enhancement of subtle spectral variations.
Hyperparameter Tuning: The best performance was achieved with $C=1$ and $\gamma=10$ . The study noted that while the default operating point did not match the specificity of frozen section (100%), adjusting the classification threshold allowed the DRS model to achieve a specificity of 0.95 while maintaining the same sensitivity as frozen section (0.85).
Prospective Validation: In the proof-of-concept testing on fresh specimens, the model successfully identified the tumor boundary with high accuracy, demonstrating the feasibility of real-time detection despite the lack of tactile feedback.

Significance and Claims
The authors claim that this study provides a prospective proof-of-concept demonstrating the feasibility of using DRS combined with machine learning to distinguish tumor from non-tumor regions in colorectal cancer surgery.

Clinical Utility: The system offers a potential solution for real-time intra-operative guidance, allowing surgeons to balance oncological safety with sphincter preservation, particularly in low rectal tumors where millimeter precision is vital.
Advantages over Current Standards: Unlike frozen sections, DRS provides near-instant results without the need for tissue removal or contrast agents. Unlike tattooing, it does not suffer from ink spreading.
Robustness: The study highlights that the model performed well even in the presence of tissue heterogeneity (e.g., mixed tumor and fibrosis following neoadjuvant therapy), a common challenge in colorectal cancer.
Future Directions: The authors modestly state that while the current results are promising, future work must focus on prospective validation in new patients during actual surgery. They also identify the need to explore probe designs with deeper penetration to better detect tumors situated deeper than the current 1–2 mm sampling depth, particularly for T2 tumors or in scenarios where the tumor extends beyond the immediate sampling zone.

The paper concludes that fast, intra-operative distal margin determination is a possibility using this technology, but emphasizes that further validation is required before clinical implementation.