Abstract
BACKGROUND: We aimed to identify the most effective machine learning model for predicting the differential diagnosis of lymph nodes (LNs) in lung cancer using dynamic and static (18)F-fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) imaging. METHODS: A total of 279 pathologically confirmed LNs from 74 patients with lung cancer were retrospectively analyzed. These were randomly divided into a training group (n = 196) and a test group (n = 83) at a ratio of 7:3. The radiomics features of the images were extracted from CT, dynamic PET (dPET), and static PET (sPET) images and were screened for the most predictive value. Support vector machine (SVM), logistic regression (LR), and random forest (RF) machine learning models were built using the optimal radiomics features. The best quantitative prediction model was suggested using SUV(max) and K (i) based on LNs. A composite model was built combining the best machine learning model and the quantitative model. Receiver operating characteristic (ROC) curves were used to evaluate the predictive ability of the machine learning, quantitative, and composite models for LN metastasis in lung cancer. RESULTS: Of the three machine learning models, the RF model demonstrated the greatest predictive efficacy in both the training [area under the curve (AUC) = 0.823] and test groups (AUC = 0.819). The quantitative model based on K (i) showed good predictive efficacy in both the training (AUC = 0.772) and test groups (AUC = 0.805). A composite model based on both the RF machine learning model and the quantitative model demonstrated superior predictive efficacy. The AUCs in the training and test groups were 0.844 and 0.835, respectively. Decision curve analysis showed that the composite model had better net benefit and clinical value. CONCLUSION: A composite model based on an RF model of PET/CT+K(i) images combined with dynamic quantitative K (i) is highly effective in differentiating FDG-avid LN metastasis in lung cancer. This model provides greater net benefit and clinical value.