Spectral CT and AI Join Forces to Predict Aggressive Bowel Cancer Before Surgery

One of the most consequential questions in colorectal cancer care is also one of the hardest to answer before surgery: how aggressive is this tumor really going to be? A new study published in BMC Medical Imaging offers a striking answer, showing that a specially engineered imaging tool can peer into the microscopic behavior of colorectal adenocarcinoma without a biopsy. The research team, led by corresponding author Kefu Liu and first authors Jinghua Chen, Lanlan Lu and Xinyu Pan, developed a clinical-radiomics nomogram built on spectral CT quantitative parameters to predict tumor budding grading preoperatively. Tumor budding, the scattering of isolated tumor cells at the invasive front of a cancer, is one of the most powerful yet hardest-to-measure indicators of a tumor’s appetite for spread. Until now, grading it required a pathologist to dissect removed tissue under a microscope, meaning patients and surgeons entered the operating room without knowing what the tumor’s biology had in store.

The implications are significant. Tumor budding grade, classified under the International Tumor Budding Consensus Conference system as low-grade (Bd1) or moderate-to-high-grade (Bd2+3), influences decisions about the extent of surgery, the urgency of lymph node evaluation, and eligibility for neoadjuvant therapy. A high-grade budding tumor behaves like a formation of soldiers abandoning ranks and infiltrating enemy territory: the individual cells detach, migrate into surrounding stroma, and seed metastasis. Knowing that risk profile before surgery could reshape treatment planning for the roughly one million people worldwide diagnosed with colorectal cancer each year. What makes the new approach remarkable is that it extracts this hidden biological information from images already acquired during routine diagnostic scans, converting ordinary CT data into a predictor of microscopic tumor behavior.

The study enrolled 138 patients with colorectal adenocarcinoma treated between May 2021 and August 2025. All patients underwent dual-layer spectral detector CT, a technology that captures images at multiple energy levels simultaneously rather than producing a single conventional picture. From each scan, the researchers measured a battery of spectral quantitative parameters: the slope of the spectral Hounsfield unit curve, effective atomic number, iodine concentration, and normalized iodine concentration, along with virtual monoenergetic images reconstructed at specific keV levels such as 40 keV. These parameters reflect the physical composition of tissue, particularly its iodine uptake, which serves as a proxy for vascularity, perfusion and cellular density. A tumor that is busily angiogenic and densely packed with cells scatters X-rays differently from a more indolent lesion, and the spectral scanner registers those differences in numbers.

Alongside imaging, the researchers collected traditional clinical variables including age, sex, carcinoembryonic antigen, carbohydrate antigen 19-9, fasting blood glucose, low-density lipoprotein cholesterol and fecal occult blood test status. Univariate analysis identified which traditional and spectral parameters differed significantly between low-grade and moderate-to-high-grade tumor budding groups. Two emerged as independent predictors: low-density lipoprotein cholesterol, a routine blood marker, and the 40 keV virtual monoenergetic image in the arterial phase. The link between LDL-C and aggressive tumor biology is intriguing, given that cholesterol metabolism has long been implicated in cancer cell membrane synthesis and proliferative signaling. Its pairing with a low-energy virtual monoenergetic image, which maximizes iodine contrast and highlights hypervascular tumor regions, suggests the model is capturing complementary windows into tumor physiology.

The radiomics arm of the study was a masterclass in modern feature engineering. From three-dimensional volumes of interest drawn on spectral quantitative parameter images, the team extracted 107 radiomics features, quantifying tumor shape, texture, intensity distribution and higher-order spatial patterns. Because moderate-to-high-grade budding was the minority class, they applied the Synthetic Minority Oversampling Technique, or SMOTE, to synthesize representative examples and mitigate class imbalance, a common pitfall that causes machine learning models to simply predict the majority class. They then ran a sequential dimensionality reduction pipeline using Pearson correlation coefficients, analysis of variance, the Relief algorithm, recursive feature elimination and the Kruskal-Wallis test. That funnel winnowed 107 features down to 23 candidates for exploratory model construction, ensuring the final model would learn genuine biological signals rather than noise or spurious correlations.

The final nomogram fused three streams of information: the radiomics score derived from spectral images, the independent clinical indicator LDL-C, and the spectral parameter from the 40 keV arterial-phase image. When validated internally, the combined clinical-radiomics nomogram delivered the strongest performance, achieving a training set area under the curve of 0.895 with a 95 percent confidence interval of 0.818 to 0.972, and a test set AUC of 0.781 with a 95 percent confidence interval of 0.688 to 0.934. In practical terms, the model correctly distinguished low-grade from moderate-to-high-grade tumor budding in roughly eight to nine out of ten training cases, and in about four out of five test cases. The team also defined an optimal classification threshold for the nomogram score, providing clinicians with a concrete decision boundary rather than an abstract probability.

Performance metrics alone do not guarantee clinical usefulness, so the researchers went further. Calibration curves demonstrated good consistency between predicted and actual outcomes across the risk spectrum, with no statistically significant divergence. Decision curve analysis, a technique that quantifies the net benefit of acting on a model’s predictions at various threshold probabilities, showed that the nomogram delivered higher clinical net benefit than either the clinical model or the radiomics model alone. This matters because a model can be statistically impressive yet practically worthless if its thresholds lead clinicians to make worse decisions than they would without it. By that standard, the combined nomogram earned its place as a genuine candidate for clinical exploration, not merely an academic exercise.

The study does carry the honest limitations of its design. It was retrospective and single-center, drawing 138 patients from one institution, and the validation was internal rather than on an external cohort. Spectral CT remains less widely deployed than conventional CT, though dual-layer detector platforms are spreading rapidly through major hospitals. The authors were also transparent that this article was shared early to provide faster access to peer-reviewed, accepted research, and the version is subject to further edits before the final Version of Record. Still, the pipeline they describe, from SMOTE balancing to multi-algorithm feature selection to nomogram construction, is readily replicable, and the modest data requirements make it feasible for other centers to test and refine the approach on their own patient populations.

If external validation confirms these findings, the consequences for colorectal cancer care could be profound. A surgeon could review a nomogram score alongside routine staging scans and know, before making an incision, whether the tumor at hand is likely to be infiltrative and bud aggressively, potentially justifying wider resection margins, more rigorous lymph node harvesting or intensified adjuvant planning. Pathologists could see the prediction corroborated by histology afterward, closing a feedback loop that continuously improves both imaging and pathology. More broadly, the study exemplifies the rise of predictive medicine, in which the boundary between imaging and molecular pathology dissolves and routine scans become windows into tumor biology. The microscope is no longer the only way to see how a cancer behaves; sometimes, it turns out, the answer has been hiding in the X-rays all along.

Subject of Research: A clinical-radiomics nomogram combining spectral CT quantitative parameters and clinical variables to predict tumor budding grading in colorectal adenocarcinoma before surgery.

Article Title: A clinical-radiomics Nomogram based on spectral CT quantitative parameters for preoperative prediction of tumor budding grading in colorectal adenocarcinoma

Article References: Chen, J., Lu, L., Pan, X., Zhu, J., Li, M., Huang, J., Tang, X., Yan, X., Qian, T., Wang, M., & Liu, K. (2026). A clinical-radiomics Nomogram based on spectral CT quantitative parameters for preoperative prediction of tumor budding grading in colorectal adenocarcinoma. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02833-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02833-y

Keywords: colorectal adenocarcinoma, spectral CT, tumor budding, radiomics, nomogram, medical imaging, machine learning, SMOTE, virtual monoenergetic images, LDL cholesterol, predictive medicine, computed tomography