Preoperative assessment of lymph node metastasis in Colon Cancer patients using machine learning: a pilot study

BACKGROUND: Preoperative detection of lymph node (LN) metastasis is critical for planning treatments in colon cancer (CC). The clinical diagnostic criteria based on the size of the LNs are not sensitive to determine metastasis using CT images. In this retrospective study, we investigated the potential value of CT texture features to diagnose LN metastasis using preoperative CT data and patient characteristics by developing quantitative prediction models. METHODS: A total of 390 CC patients, undergone surgical resection, were enrolled in this monocentric study. 390 histologically validated LNs were collected from patients and randomly separated into training (312 patients, 155 metastatic and 157 normal LNs) and test cohorts (78 patients, 39 metastatic and 39 normal LNs). Six patient characteristics and 146 quantitative CT imaging features were analyzed and key variables were determined using either exhaustive search or least absolute shrinkage algorithm. Two kernel-based support vector machine classifiers (patient-characteristic model and radiomic-derived model), generated with 10-fold cross-validation, were compared with the clinical model that utilizes long-axis diameter for diagnosis of metastatic LN. The performance of the models was evaluated on the test cohort by computing accuracy, sensitivity, specificity, and area under the receiver operating curve (AUC). RESULTS: The clinical model had an overall diagnostic accuracy of 64.87%; specifically, accuracy of 65.38% and 62.82%, sensitivity of 83.87% and 84.62%, and specificity of 47.13% and 41.03% for training and test cohorts, respectively. The patient-demographic model obtained accuracy of 67.31% and 73.08%, the sensitivity of 62.58% and 69.23%, and specificity of 71.97% and 76.23% for training and test cohorts, respectively. Besides, the radiomic-derived model resulted in an accuracy of 81.09% and 79.49%, sensitivity of 83.87% and 74.36%, and specificity of 78.34% and 84.62% for training and test cohorts, respectively. Furthermore, the diagnostic performance of the radiomic-derived model was significantly higher than clinical and patient-demographic models (p < 0.02) according to the DeLong method. CONCLUSIONS: The texture of the LNs provided characteristic information about the histological status of the LNs. The radiomic-derived model leveraging LN texture provides better preoperative diagnostic accuracy for the detection of metastatic LNs compared to the clinically accepted diagnostic criteria and patient-demographic model.