Preoperative prediction of mesenteric lymph node metastasis in colorectal cancer using machine learning with CT-based data

Colorectal cancer is one of the most prevalent malignancies globally, ranking third in incidence and second in cancer-related mortality. Previous research has emphasized the lymph node as a key metastatic pathway in colorectal cancer, with more than 53% of patients showing mesenteric lymph node involvement associated with the tumor. As a result, the detection of metastasis in lymph nodes has emerged as a crucial prognostic factor for patients battling colorectal cancer, providing essential information for clinical decision-making. Curative-intent surgery, especially radical lymphadenectomy, remains the primary treatment for non-metastatic colorectal cancer. While this approach can be life-saving, it also carries the risk of overtreatment, potentially leading to unnecessary surgical interventions in patients who may not benefit from such aggressive measures. Furthermore, mesenteric lymph node metastasis is also essential for prognosis and identifying patients who are likely to derive benefit from adjuvant therapy. Currently, lymph node involvement is evaluated through N staging, which relies on histological examination of lymph node specimens to determine metastatic progression. However, there is a pressing need for more expedient and accurate methods that can predict mesenteric lymph node metastasis preoperatively.

Recently, a research team led by Lanni Zhou, Fusheng Ouyang, Qiugen Hu and colleagues from the Department of Radiology at Shunde Hospital, Southern Medical University, China, published a pivotal study titled "Enhanced prediction of preoperative mesenteric lymph node metastasis in colorectal cancer using machine learning with CT-based data " in MedComm-Future Medicine, a journal published by Wiley. This study aims to develop a clinical prediction model using machine learning algorithms to assess the risk of mesenteric lymph node metastasis preoperatively, based on CT images and clinicopathological data from colorectal cancer patients.

Medical records were collected from the hospital’s electronic medical record system. Among 147 patients, 49 (33.33%) had mesenteric lymph node metastasis, while 98 (66.67%) did not. The group with mesenteric lymph node metastasis exhibited significantly higher levels of maximum tumor diameter, platelet counts, low differentiation degree ratio, lymphovascular invasion rate, and perineural invasion rate compared to the non- mesenteric lymph node metastasis group. Conversely, the mesenteric lymph node metastasis group demonstrated notably lower hematocrit levels. The researchers employed least absolute shrinkage and selection operator regression analysis to analyze these findings, using the presence of mesenteric lymph node metastasis as the dependent variable. Finally, seven variables were screened out, including perineural invasion, lymphovascular invasion, extracellular volume fraction of tumor, the difference in CT attenuation value between the equilibrium phase and no-contrast phase (ΔHUtumor), platelet to lymphocyte ratio, hematocrit level and differentiation degree.

The study used five widely used and current machine learning techniques to construct the predictive models, including Random Forest, Naive Bayes, Extreme Gradient Boosting (XGB), Light Gradient Boosting, and K-Nearest Neighbor models, with each model undergoing 10 repetitions to ensure robustness. Among the five models, the XGB model demonstrated the highest level of stability, achieving an AUC of 0.836 in the training group and 0.831 in the validation group. Notably, the XGB model exhibited a smaller standard deviation in the AUC score for the validation set (0.059), outperforming the other models in terms of consistency. Moreover, the predictive performance of the XGB model is superior to that of each single factor in the model. These findings suggest that the XGB model exhibited the most reliable performance. The model was further elucidated using SHapley Additive Explanation values, which ranked predictors in the XGB model by their importance, providing valuable insights into the decision-making process of the model. Calibration and decision curve analyses showed good calibration and clinical benefit, underscoring the model's applicability in a clinical setting.

In summary, this study successfully established an XGB-based machine learning model to predict the likelihood of mesenteric lymph node metastasis in patients with colorectal cancer. The resulting model shows significant potential to assist in clinical treatment planning, optimize the selection of surgical methods, and guide decision-making for adjuvant therapy prior to surgery.

See the article: 

Enhanced prediction of preoperative mesenteric lymph node metastasis in colorectal cancer using machine learning with CT-based data