# Machine Learning Radiomics in Computed Tomography for Prediction of Tumor and Nodal Stages in Colorectal Cancer

## TL;DR

This study shows that machine learning and radiomics from CT scans can help predict the stage of colorectal cancer before surgery, improving treatment planning.

## Contribution

The study introduces a noninvasive CT-based radiomics framework for preoperative T and N staging in colorectal cancer using machine learning.

## Key findings

- Shape-based features improved tumor staging with a balanced accuracy of 0.70 and AUC of 0.751.
- Texture-based features enhanced nodal staging with a sensitivity of 0.742 and AUC of 0.750.
- Radiomics captured morphological and textural patterns linked to tumor invasiveness and lymphatic spread.

## Abstract

Colorectal cancer is a common disease in which treatment decisions depend greatly on how deeply the tumor grows and whether it has already spread to nearby lymph nodes. However, standard scans often struggle to clearly show these details before surgery. Radiomics is a new imaging approach that extracts many invisible patterns from routine scans to better understand tumor behavior. In this study, we analyzed computed tomography images from patients with colorectal cancer and used computer-based models to determine whether radiomic patterns could help identify early or advanced tumors and predict lymph node involvement. We found that specific image features related to tumor shape and texture provided valuable information for both tasks. These findings suggest that radiomics could become a useful, noninvasive tool to improve preoperative evaluation and support more personalized treatment planning in the future.

Background/Objectives: Accurate preoperative TN staging is essential for guiding surgical and adjuvant treatment decisions in colorectal cancer (CRC), yet conventional imaging still faces limitations in reliably distinguishing early from advanced disease. This study aimed to evaluate whether CT-based radiomics combined with machine learning can noninvasively predict both tumor (T) and nodal (N) stages of CRC, and to identify which feature groups most contribute to each task. Methods: Fifty-three patients (55 tumors) with histologically confirmed CRC who underwent preoperative contrast-enhanced CT were retrospectively analyzed. A total of 107 radiomic features were extracted using PyRadiomics version 3.1.0, including shape, first-order, and texture features. Multiple preprocessing strategies—z-score normalization, PCA, and SMOTE—were tested across 11 machine learning classifiers. Results: For T staging, logistic regression using shape-based features achieved a mean sensitivity of 0.721, a specificity of 0.68, a balanced accuracy of 0.70, and an AUC of 0.751. For N staging, the AdaBoost model using texture-based features achieved a sensitivity of 0.742, a specificity of 0.622, a balanced accuracy of 0.682, and an AUC of 0.750. Shape features predominantly contributed to T prediction, while texture matrices drove N prediction, reflecting morphological and microstructural correlates of invasiveness and lymphatic dissemination. Conclusions: CT-based radiomics can quantitatively capture both morphological and textural patterns of tumor behavior, providing a noninvasive framework for preoperative TN staging in CRC.