Development and validation of machine learning models for predicting no. 253 lymph node metastasis in left-sided colorectal cancer using clinical and CT-based radiomic features

Abstract

Background: The appropriate ligation level of the inferior mesenteric artery (IMA) in left-sided colorectal cancer (CRC) surgery is debated, with metastasis in No. 253 lymph node (No. 253 LN) being a key determining factor. This study aimed to develop a machine learning model for predicting metastasis in No. 253 LN.

Methods: We retrospectively collected clinical data from 2,118 patients with left-sided CRC and contrast-enhanced CT images from 310 of these patients. From this data, a test set, a training set, and a temporal validation set were constructed. Logistic regression models were used to develop a clinical model, a CT model, and a radiomics model, which were then integrated into a combined model using logical rules. Finally, these models were evaluated using metrics such as the area under the receiver operating characteristic curve (AUC), precision-recall (PR) curves, decision curve analysis (DCA), net reclassification improvement (NRI), and integrated discrimination improvement (IDI).

Results: A clinical model, a CT model, and a radiomics model were constructed using univariate logistic regression. A combined model was developed by integrating the clinical, CT, and radiomics models, with positivity defined as all three models being positive at a 90% sensitivity threshold. The clinical model included six predictive factors: tumor site, endoscopic obstruction, CEA levels, growth type, differentiation grade, and pathological classification. The CT model utilized largest lymph node average CT value, short-axis diameter and long-axis diameter. The radiomics model incorporated maximum gray level intensity within the region of interest, large area high gray level emphasis, small area high gray level emphasis and surface area to volume ratio. In the test set, the AUCs for the clinical, CT, radiomics, and combined models were 0.694, 0.663, 0.72, and 0.663, respectively, while in the temporal validation set, they were 0.743, 0.629, 0.716, and 0.8. Specifically, the combined model demonstrated a sensitivity of 0.8 and a specificity of 0.8 in the temporal validation set. By comparing the PR and DCA curves, the combined model demonstrated better performance. Additionally, the combined model showed moderate improvements in INR and IDI compared to other models.

Conclusion: A clinical and CT-based radiomics model shows promise in predicting No. 253 LN metastasis in left-sided CRC and provides insights for optimizing IMA ligation strategies.

Keywords: Colorectal cancer; Lymph node metastasis; Machine learning; Prediction model; Radiomics.

Fig. 1

The annotation of No. 253 LN region on CT images (a). The four views represent axial (top left), coronal (bottom left), sagittal (bottom right), and 3D (top right) images. The red area represents the aorta, the green area represents IMV, the blue area represents IMA, and the yellow area represents LCA. The upper boundary for annotating No. 253 LN was the starting point of IMA, with the right boundary being the left edge of the aorta, the left boundary being the right edge of IMV, and the lower boundary extending from IMA to the intersection of LCA and IMV. The annotation of all the lymph nodes in No. 253 LN region (b). The purple area represents the largest lymph node and the light blue area represents other lymph nodes in No. 253 LN region. LN = lymph node; IMV = inferior mesenteric vein; IMA = inferior mesenteric artery; LCA = left colic artery

Fig. 2

The construction methodology of clinical, CT, Radiomics, and combined model. A large-scale clinical cohort was used to train the clinical model, while imaging data were collected from high-risk individuals identified through propensity score matching based on clinical information to construct an imaging cohort for training the CT and radiomics models. Then, a clinical model, a CT model, and a radiomics model were constructed using univariate logistic regression. Subsequently, a combined model was developed by integrating the clinical, CT, and radiomics models, with positivity defined as all three models being positive at a 90% sensitivity threshold. Finally, the models were evaluated using a temporal validation cohort. PSM = propensity score matching

Fig. 3

Study flow diagram. LN = lymph node

Fig. 4

Comparison of the clinical, CT, radiomics, and combined models. ROC curves in the test set (a). ROC curves in the temporal validation set (b). PR curves in the test set (c). PR curves in the temporal validation set (d). DCA curves in the test set (e). DCA curves in the temporal validation set (f). ROC = receiver operating characteristic; PR = precision recall