Prediction of mediastinal lymph node metastasis based on ¹⁸F-FDG PET/CT imaging using support vector machine in non-small cell lung cancer

Objective: The purpose of this study was to develop a classification method based on support vector machine (SVM) to improve the diagnostic performance of ¹⁸F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) to detect the lymph node (LN) metastasis in non-small cell lung cancer (NSCLC).

Method: Two hundred nineteen lymph nodes (37 metastatic) from 71 patients were evaluated in this study. SVM models were developed with 7 LN features. The area under the curve (AUC) and accuracy of 9 models were compared to select the best model. The best SVM model was simplified on the basis of the feature weights and value distribution to further suit the clinical application.

Results: The maximum, minimum, and mean accuracy of the best model was 91.89% (68/74, 95% CI 83.11~96.54%), 66.22% (49/74, 95% CI 54.85~75.98%), and 80.09% (59,266/74,000, 95% CI 70.27~89.19%), respectively, with an AUC of 0.94, 0.66, and 0.81, respectively. The best SVM model was finally simplified into a score rule: LNs with scores more than 3.0 were considered as malignant ones, whereas LNs with scores less than 1.5 tended to be benign ones. For the LNs with scores within a range of 1.5-3.0, metastasis was suspected.

Conclusion: An SVM model based on ¹⁸F-FDG PET/CT images was able to predict the metastatic LNs for patients with NSCLC. The ratio of the maximum of standard uptake value of LNs to aortic arch played a major role in the model. After simplification, the model could be transferred into a scoring method which may partly help clinicians determine the clinical staging of patients with NSCLC relatively easier.

Key points: • The SVM model based on ¹⁸F-FDG PET/CT features may help clinicians to make a decision for metastatic mediastinal lymph nodes in patients with NSCLC. • The SUR_blood plays a major role in the SVM model. • The score rule based on the SVM model simplified the complexity of the model and may partly help clinicians determine the clinical staging of patients with NSCLC relatively easier.