F-18 FDG PET/CT based Preoperative Machine Learning Prediction Models for Evaluating Regional Lymph Node Metastasis Status of Patients with Colon Cancer.

OBJECTIVE

This study aimed to develop a simple machine-learning model incorporating lymph node metastasis status with F-18 Fluorodeoxyglucose positron emission tomography/computed tomography (FDG PET/CT) and clinical information for predicting regional lymph node metastasis in patients with colon cancer.

METHODS

This retrospective study included 193 patients diagnosed with colon cancer between January 2014 and December 2017. All patients underwent F-18 FDG PET/CT and blood test before surgery. One categorical variable (lymph node FDG uptake [LNFDG]) and six continuous variables (age, neutrophil-to-lymphocyte ratio [NLR], carcinoembryonic antigen [CEA], carbohydrate antigen 19-9 [CA19-9], C-reactive protein, and maximal standardized uptake value (SUVmax) of the primary tumor) were used as input variables. Four supervised machine learning methods were used to build predictive models: logistic regression (LR), random forest (RF), gradient boosting machine (GBM), and support vector machine (SVM). Area under the receiver operating characteristic curve (AUC) of the validation set were used for evaluating and comparing model performance.

RESULTS

The number of patients with lymph node metastasis were 63 (33%). The mean number of harvested lymph nodes was 28.8 ± 11.4. The mean CEA, CA19-9, and CRP levels were 4.8 ± 9.3 ng/ml, 15.6 ± 42.8 U/ml, and 1.0 ± 3.0 mg/dl, respectively. The mean NLR was 2.2 ± 1.2. The mean SUVmax levels of the primary tumor were 15.2 ± 7.9. Fifty-one (26%) patients showed FDG uptake in the pericolic lymph nodes.  The mean AUC of the LR, RF, GBM, and SVM methods for the LNFDG model was 0.7046, 0.7047, 0.7040, and 0.7058, respectively. The mean AUC of the LR, RF, GBM, and SVM methods for the LNFDG plus clinical information model was 0.7046, 0.7302, 0.7444, and 0.7097, respectively.

CONCLUSION

Machine learning methods using LNFDG and clinical information could predict the lymph node metastasis status in patients with colon cancer with higher accuracy than a model using only FDG uptake of the lymph nodes.