Carcinoembryonic antigen (CEA) is a protein that can be measured in the blood and is one of the most commonly used biomarkers in colorectal cancer management. It helps doctors monitor patients after treatment and can suggest disease recurrence when levels rise over time. In healthy non-smoking adults, CEA levels below 3.0 nanograms per milliliter are considered normal, while in smokers the upper limit rises to 5 ng/mL, which is also the threshold commonly used in clinical practice to define an elevated result.
CEA levels between 5 and 10 ng/mL are generally associated with localized disease and a relatively favorable prognosis, while levels exceeding 10 ng/mL indicate a higher risk of recurrence and poorer outcomes. However, CEA has important limitations - it is not always proportional to tumor size or aggressiveness, and some patients with advanced colorectal cancer have perfectly normal CEA levels, while others with elevated CEA may not have active cancer. This biological variability leads to false-negative and false-positive results.
Because of these limitations, researchers have been searching for complementary tools that can provide additional information about tumor biology. The ideal complement would be non-invasive, objective, and capable of revealing characteristics of the tumor that a blood test alone cannot capture. Radiomics - the extraction of quantitative features from medical images - has emerged as a promising approach for this purpose.
Radiomics is a process of transforming standard medical images into large sets of quantitative data. Rather than simply looking at an image to determine whether a tumor is present, radiomics extracts measurable numerical features from every pixel and region of interest in the image - features describing intensity, shape, and texture at a level of detail far beyond what the human eye can perceive. Hundreds of such features can be derived from a single CT scan.
These features are thought to reflect the underlying biological properties of tumors, including tumor heterogeneity (variation within the tumor), angiogenesis (the formation of new blood vessels), and even gene expression patterns. The key insight driving radiomics research is that the microscopic biology of a tumor leaves a detectable fingerprint in the way it appears on imaging - a fingerprint that can be decoded with the right computational tools.
In colorectal cancer, previous studies have shown that CT-based radiomic features can predict survival, the risk of metastasis, and treatment response. This study took the next step: investigating whether radiomic features could also predict CEA status - effectively asking whether a CT scan can provide the same information as a blood test, and whether combining both might offer better clinical decision support.
This was a retrospective study conducted at a single institution in Turkey. Researchers initially screened 150 adults with histopathologically confirmed colon cancer who had undergone pre-operative CT imaging and had serum CEA measurements available within 10 days of their scan. After excluding 41 patients due to poor image quality or incomplete data, the final cohort consisted of 109 eligible patients.
Using the clinical threshold of 5 ng/mL, patients were divided into two groups: 53 CEA-positive patients (with CEA at or above 5 ng/mL) and 56 CEA-negative patients (with CEA below 5 ng/mL). The groups were well matched in size, which is important for training fair and unbiased machine learning models. The cohort had a mean age of 55.09 years, and 64.2% were male.
The study population was diagnostically diverse, including patients across multiple TNM stages - from early stage I disease to advanced stage IV. The most common tumor type was adenocarcinoma, representing 96.3% of cases. Importantly, 34.9% of patients were smokers, and this smoking status turned out to be a significant factor in the analysis, as discussed in the results.
All CT examinations were performed using a 64-slice CT scanner with a standardized protocol: 120 kV tube voltage, contrast agent administered intravenously at 1 mL per kilogram body weight, and imaging acquired 65 seconds after injection in the portal venous phase. This phase provides optimal contrast between colorectal tumors and surrounding normal tissue, maximizing the visibility of tumor boundaries.
Tumor regions were manually segmented (outlined) by an experienced radiologist with 13 years of dedicated abdominal imaging expertise, using the open-source software 3D Slicer. For each patient, three image slices were analyzed - the central slice showing the largest tumor cross-section, plus the immediately adjacent slices above and below. This approach captures three-dimensional spatial information while keeping the segmentation task manageable.
Following segmentation, the PyRadiomics library was used to extract 107 quantitative features from each tumor region. These features spanned three main categories: first-order statistics (describing the overall distribution of pixel intensities), shape features (describing the geometric form of the tumor), and texture features derived from mathematical models including the Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run-Length Matrix (GLRLM), which capture fine-grained patterns of spatial variation within the tumor.
Before training any model, the researchers performed two quality-control steps on the radiomic features. First, reproducibility testing: features were extracted twice by the same radiologist two weeks apart (intra-observer) and once by a second radiologist (inter-observer). Only features with an Intraclass Correlation Coefficient (ICC) above 0.8 - indicating strong consistency between repeated measurements - were retained.
Second, correlation-based feature selection was applied to eliminate redundant features. When two features are very highly correlated with each other (|r| >= 0.8), they carry essentially duplicate information, and keeping both wastes computational resources while risking overfitting. The feature with lower variance in each correlated pair was removed. This process reduced the feature set from 107 to a final set of 41 robust, non-redundant features.
Five different machine learning classifiers were then trained and compared: Support Vector Machine (SVM), which finds an optimal boundary between classes; Decision Tree, which makes sequential yes/no decisions; Ensemble Learning, which combines multiple weak learners; k-Nearest Neighbor (k-NN), which classifies new samples based on the most similar training examples; and Neural Network (NN), which uses layered computational nodes to model complex patterns. All models were evaluated using stratified 5-fold cross-validation, where the data was split five times to ensure each patient served as a test case at least once.
Using the final set of 41 selected features, the k-Nearest Neighbor (k-NN) classifier achieved the highest overall performance: accuracy of 77.4% and ROC-AUC of 0.8523. An AUC (Area Under the Receiver Operating Characteristic Curve) value above 0.85 is generally considered good discrimination performance, indicating that the model can meaningfully distinguish CEA-positive from CEA-negative patients based on their CT features.
The Support Vector Machine (SVM) and Neural Network (NN) models achieved the highest recall (sensitivity) of 83.0%, meaning they correctly identified 83% of CEA-positive patients. High recall is particularly important in clinical settings where missing a positive case could have serious consequences. Different classifiers thus showed complementary strengths: k-NN excelled in overall accuracy and AUC, while SVM and NN were better at detecting true positives.
Performance across classifiers was broadly similar, suggesting that the radiomic features themselves carried meaningful discriminative information regardless of which model was used to analyze them. The fact that multiple independent algorithms converged on similar results strengthens confidence that the radiomic signal is real rather than being an artifact of one particular modeling approach.
An important ancillary finding of this study was the strong association between smoking status and serum CEA levels. Smokers in the cohort had a mean CEA of 20.78 ng/mL, compared to only 4.08 ng/mL in non-smokers - a more than five-fold difference. This difference was statistically highly significant (p = 0.001) and remained significant after adjusting for other clinical variables in a multivariate analysis.
In the multivariate linear regression model, smoking was the only variable that remained significantly associated with elevated CEA (beta = 28.589, p = 0.003). Age, sex, tumor stage, and tumor type were all non-significant. This finding is consistent with the well-known biological fact that tobacco compounds stimulate CEA production independently of cancer, which is why the clinical CEA threshold is higher for smokers (5 ng/mL) than for non-smokers (3 ng/mL).
This result has practical implications for interpreting CEA results and for training models that use CEA as a label. It suggests that smoking status should be considered a confounding factor in any analysis involving CEA - both in clinical interpretation and in machine learning studies - and that models attempting to predict CEA from imaging features might perform differently in smokers versus non-smokers.
The study demonstrates that CT-based radiomic features combined with machine learning can predict CEA status with accuracy and AUC values that suggest genuine clinical utility. The ability to derive CEA-related information from imaging - without requiring a blood draw or waiting for laboratory results - could be valuable in situations where blood test timing or availability is limited, or as a complementary tool to refine risk assessment.
More broadly, these findings support the concept that imaging biomarkers can reflect the molecular biology of tumors. If radiomic features can distinguish CEA-positive from CEA-negative tumors, they may also carry information about tumor aggressiveness, treatment response, and prognosis that goes beyond what is visible to the human eye. This positions radiomics as a non-invasive tissue characterization tool that could enhance personalized treatment decisions.
Important limitations temper these conclusions. The study enrolled only 109 patients from a single institution, which is relatively small for machine learning research and limits statistical power and generalizability. The retrospective design introduces the potential for selection bias. External validation in independent patient cohorts from different institutions is needed before these models can be considered for clinical implementation. Future research should also explore which specific radiomic features are most biologically meaningful and investigate whether combining radiomic and genomic data improves performance further.