# Radiomics profiling combined with clinical risk factors for preoperative Lymphatic Metastasis prediction in Colorectal cancer: A multicenter study

## TL;DR

This study combines radiomics and clinical data to improve preoperative prediction of lymphatic metastasis in colorectal cancer patients.

## Contribution

A novel machine learning model integrating radiomic features and clinical risk factors for accurate preoperative lymphatic metastasis prediction in CRC.

## Key findings

- The combined model (ModelC_3D(R + C)) outperformed other models with an AUC of 0.858 in training and 0.833 in validation.
- Incorporating clinical and radiomic features significantly improved prediction accuracy compared to models using only clinical or radiomic data.
- The model demonstrated superior discriminative performance in identifying lymphatic metastasis in colorectal cancer patients.

## Abstract

Accurate preoperative assessment of regional lymphatic metastases (LNM) is essential for effective surgical selection of patients with colorectal cancer (CRC). This study aimed to develop a machine learning (ML) model that integrates radiomics and clinical risk factors to predict preoperative LNM in CRC patients.

This multicenter cohort study retrospectively collected data from 349 CRC patients between January 1, 2020, and December 31, 2023. A total of 292 patients from our hospital comprised the training dataset, while 57 patients from external hospitals formed the validation dataset. Radiomic features of the tumor region (3D(R)) and colorectal region (3D(C)) were extracted from venous-phase CT images. LASSO (least absolute shrinkage and selection operator) regression was applied to screen clinical and radiomic features. 4 prediction models, clinical, 3D(R), 3D(R + C),and combined, were constructed using support vector machine (SVM). The optimal model was identified through comparative analysis of the area under the curve (AUC) metric across multiple models.

The Model_3D(R + C) demonstrated superior discriminative performance compared to Model_3D(R) alone (AUC: training, 0.733(95% CI: [0.693, 0.773]) vs. 0.696 (95% CI: [0.655, 0.737]); validation, 0.641(95% CI: [0.590, 0.692]) vs. 0.563(95% CI: [0.507, 0.619])). The model combining clinical and 3D(R + C) (ModelC_3D(R + C))outperformed the clinical model(ModelC) and Model_3D(R + C) (AUC: training: 0.858(95% CI: [0.826, 0.890]) vs. 0.635(95% CI: [0.585, 0.685]) vs. 0.733(95% CI: [0.693, 0.773]); validation 0.833(95% CI: [0.787, 0.879]) vs. 0.589(95% CI: [0.537, 0.641]) vs. 0.641(95% CI: [0.590, 0.692]); P < 0.050). Therefore, the combined model provided the most accurate identification of LNM.

The SVM model incorporating 3D(R) features, 3D(C) features, and clinical risk factors effectively predicts preoperative LNM in CRC patients.