📄 medicine

Intratumoral and multiscale peritumoral radiomics on non-contrast CT for differentiating rectal adenocarcinoma from rectal neuroendocrine tumors: a retrospective diagnostic study

This retrospective study demonstrates that a machine learning model integrating intratumoral and multiscale peritumoral radiomics features from non-contrast CT with clinical variables achieves the highest point estimate for differentiating rectal adenocarcinoma from rectal neuroendocrine tumors, though the findings remain exploratory due to the small, single-center cohort and lack of external validation.

Original authors: Ximing Liu¹, Yujie Zhou¹, Lisong Zhu², Shuang Gao³, Renchen Dou³, Rixiong Wang¹

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ximing Liu¹, Yujie Zhou¹, Lisong Zhu², Shuang Gao³, Renchen Dou³, Rixiong Wang¹

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

In the complex landscape of the human body, the rectum can harbor two very different kinds of tumors that look similar on a scan but require completely different treatments. One is a common form of cancer that grows aggressively, while the other is a slower-growing tumor arising from nerve-like cells. Distinguishing between them before surgery is crucial, as it dictates whether a patient needs a radical operation or a simpler procedure. Usually, doctors rely on contrast-enhanced scans, where an injected dye highlights blood vessels to reveal the tumor's nature. However, not every patient can receive this dye due to allergies or kidney issues, and sometimes only a plain scan is available. This leaves a diagnostic gap where doctors must look at the tumor's shape and texture without the help of color-enhanced clues.

A team of researchers set out to see if they could fill this gap using a method called radiomics. Instead of asking a human eye to guess the difference, they treated the medical images as a vast source of data. They broke down the pictures of the tumors into thousands of tiny numerical details—measuring the grayness, the roughness, and the patterns of the pixels that are too subtle for a person to see. They also looked not just at the tumor itself, but at the tissue immediately surrounding it, reasoning that the tumor's influence might ripple out into its neighbors. By feeding these hidden patterns into a computer program capable of learning from examples, they hoped to build a tool that could tell the two types of tumors apart using only the plain, non-contrast images.

The study focused on 87 patients who had been treated for rectal tumors between 2019 and 2025. The researchers divided these cases into two groups: one to teach the computer what to look for, and another to test if the computer could apply what it learned to new, unseen cases. They carefully drew outlines around the tumors on the plain CT scans and then expanded those outlines into the surrounding tissue by 3 millimeters and 5 millimeters, creating three distinct zones to analyze. From these zones, they extracted over a thousand features for each patient, such as the texture of the tumor's surface and the distribution of gray shades in the nearby tissue. To ensure these measurements were reliable, they checked that different observers would draw the same lines and get the same numbers, keeping only the most consistent data.

The computer was then trained to find the specific combination of these tiny details that best separated the aggressive adenocarcinomas from the neuroendocrine tumors. The researchers tested several different mathematical approaches, but one method, known as extreme gradient boosting, proved most effective. They found that looking at the tissue just 3 millimeters outside the tumor boundary provided the most useful information, even more so than looking at the tumor itself or the tissue 5 millimeters away. This suggests that the immediate neighborhood of the tumor holds a unique signature that helps identify its type. When the researchers combined this imaging data with standard clinical information, such as the patient's age and the size of the tumor, the computer's ability to distinguish between the two conditions improved further.

The results showed that this combined approach achieved a high discrimination score of 0.879 on a scale where 1.0 represents perfect ranking ability, outperforming models using only clinical data or tumor imaging alone. However, the researchers were careful to note that while the model ranked cases well overall, it struggled when forced to make a final yes-or-no decision at a specific cutoff point. In the test group, the model failed to identify 70% of the neuroendocrine tumors when using the standard threshold derived from the training data. This high false-negative rate indicates that the model was not yet sensitive enough to be used as a standalone diagnostic tool, despite its strong overall ranking performance. The study also highlighted that the features the computer found most important were not the ones that had been statistically proven to be different in a traditional sense, but rather a complex mix of patterns that only emerged when analyzed together.

Ultimately, this work serves as a proof of concept rather than a finished medical product. It demonstrates that valuable diagnostic clues exist within plain CT scans, hidden in the texture of the tumor and the tissue around it, waiting to be unlocked by advanced computing. The findings suggest that a future tool could help doctors make better decisions when contrast dye is not an option. Yet, because the study was conducted at a single hospital with a relatively small number of patients, the results are considered exploratory. The researchers emphasize that before such a system could be used in real clinics, it must be tested on a much larger and more diverse group of people to ensure it works reliably for everyone. For now, it offers a promising glimpse into how digital analysis might one day support human judgment in the most difficult diagnostic scenarios.

Technical Summary: Intratumoral and Multiscale Peritumoral Radiomics on Non-Contrast CT for Differentiating Rectal Adenocarcinoma from Rectal Neuroendocrine Tumors

Problem Statement
Distinguishing rectal adenocarcinoma (RAC) from rectal neuroendocrine tumors (RNETs) prior to treatment is clinically critical due to divergent management strategies. While contrast-enhanced CT can reveal vascular clues characteristic of RNETs, iodinated contrast is often contraindicated or unavailable. Consequently, there is a need to determine whether quantitative features derived from non-contrast CT—specifically intratumoral and peritumoral radiomics—can effectively discriminate between these two entities without relying on enhancement data.

Methodology
This single-center, retrospective diagnostic study analyzed 87 pathologically confirmed cases (67 RAC, 20 RNET) treated between January 2019 and December 2025. The cohort was split approximately 1:1 into a training set ( $n=44$ ) and an internal validation set ( $n=43$ ).

Data Acquisition and Preprocessing: Analyses were restricted to pretreatment non-contrast CT images. Images from various scanners were resampled to isotropic $1\times1\times1$ mm³ voxels, and gray-level discretization was performed with a fixed bin width of 25.
Region of Interest (ROI) Segmentation: A radiologist manually delineated intratumoral ROIs, which were then expanded to create 3-mm and 5-mm peritumoral regions. Inter- and intra-observer reproducibility was assessed using the intraclass correlation coefficient (ICC), with a threshold of $\ge0.80$ used to filter stable features.
Feature Extraction and Selection: A total of 1,561 radiomics features were extracted per region using PyRadiomics. Feature selection followed a rigorous pipeline restricted to the training cohort:
1. Reproducibility filtering (ICC $\ge0.80$ ).
2. Univariable statistical testing ( $P<0.05$ ).
3. Redundancy removal via Spearman correlation ( $|\rho| \ge 0.90$ ).
4. Ranking via Maximum Relevance Minimum Redundancy (mRMR).
5. Final selection via LASSO regression to generate regional Radiomics Scores (RadScores).
Clinical Variables: Candidate clinical variables included age, sex, T/N/M stages, metastasis status, tumor size, and tumor markers (AFP, CEA, CA125, CA19-9). Selection utilized Firth logistic regression and LASSO.
Model Development: Five classifiers (LR, RF, XGB, SVM, KNN) were evaluated. Nine feature configurations were tested: clinical-only, single-region radiomics (intratumoral, 3-mm, 5-mm), combined radiomics, and clinical-radiomics fusion models. Extreme Gradient Boosting (XGB) was selected as the representative algorithm due to consistent performance.
Evaluation: Models were assessed using AUROC, calibration curves, Decision Curve Analysis (DCA), and DeLong testing for pairwise comparisons. Model interpretability was explored using SHAP (Shapley Additive Explanations) based on probability surrogates.

Key Results

Feature Stability: After ICC filtering, 1,240 (intratumoral), 1,074 (3-mm peritumoral), and 1,130 (5-mm peritumoral) stable features were retained. Final LASSO selection yielded eight features for each region.
Discriminative Performance:
- Among single-region models, the 3-mm peritumoral XGB model achieved the highest internal-validation AUROC of 0.762 (95% CI: 0.595–0.911), outperforming the intratumoral (0.732) and 5-mm peritumoral (0.717) models.
- The clinical-radiomics fusion XGB model (combining clinical variables with all three RadScores) yielded the highest point estimate for internal-validation AUROC at 0.879 (95% CI: 0.721–0.982).
- Pairwise DeLong comparisons between the clinical, radiomics, and fusion models were not statistically significant ( $P > 0.05$ ) in either cohort.
Threshold Sensitivity: At the Youden threshold derived from the training cohort, the fusion model exhibited a false-negative rate (FNR) of 0.700 in the internal validation set, indicating that high AUROC did not translate to high sensitivity at this specific operating point.
Feature Importance: SHAP analysis revealed that the 3-mm peritumoral RadScore was the most influential variable in the fusion model, followed by the 5-mm peritumoral RadScore and T stage.

Significance and Claims
The authors frame this work as an exploratory, hypothesis-generating study rather than a confirmation of clinical generalizability. The paper claims that:

Peritumoral Information is Valuable: Non-contrast CT peritumoral radiomics, particularly at a 3-mm expansion, provides complementary discriminatory information that may capture tumor-host interface characteristics (e.g., stromal reaction, microvasculature) not visible in the intratumoral ROI alone.
Feasibility of Fusion Models: Integrating clinical variables with multiscale radiomics yields the highest internal-validation performance point estimate, suggesting that imaging phenotypes and clinical disease extent offer complementary signals.
Limitations and Future Needs: The authors explicitly state that the findings are limited by the small, single-center cohort, the lack of external validation, and the fact that selected features did not survive False Discovery Rate (FDR) correction. They emphasize that the model is not ready for clinical deployment. Before clinical application, a "locked" model requires external validation in an independent multicenter cohort, and operating thresholds must be recalibrated to address the high false-negative rate observed at the training-derived cutoff.

In summary, the study demonstrates the technical feasibility of using non-contrast CT radiomics to differentiate RAC from RNETs, highlighting the specific utility of the 3-mm peritumoral region, while maintaining a cautious stance regarding its current clinical utility.

Technical Summary: Intratumoral and Multiscale Peritumoral Radiomics on Non-Contrast CT for Differentiating Rectal Adenocarcinoma from Rectal Neuroendocrine Tumors

More like this