Purpose: Since the COVID-19 mRNA vaccine rollout, there have been reports of FDG uptake in the draining axillary lymph nodes (ALNs) of the injection side in PET/CT. We evaluated if radiomics with machine learning can differentiate between FDG-avid metastatic axillary lymphadenopathy and FDG-avid reactive axillary lymph nodes.
Methods: We retrospectively analyzed FDG-positive, pathology-proven, ALNs in 53 patients with breast cancer who had PET/CT for follow-up or staging and FDG-positive ALNs in 46 patients who were vaccinated with the COVID-19 mRNA vaccine and had PET/CT for various reasons. Each LN was segmented in the PET and CT images. Radiomics features were extracted from all segmented LNs. Analysis was performed on PET, CT, and combined PET/CT inputs. K-nearest neighbors (KNN) and random forest (RF) machine learning models were used to train and test the radiomics features and predict the best classification between cohorts. Performance was evaluated by AUC score.
Results: A total of 165 segmented ALNs were included: 85 from patients with breast cancer and 80 from COVID-19-vaccinated individuals. Analysis of the first-order features group showed statistically significant differences (p<0.05) in all combined PET/CT radiomics features, in most PET features, and in half of the CT tested features. Combined PET/CT input using KNN model showed the best performance score, with 0.98±0.03 validation AUC and 96%±4% validation accuracy. The RF model showed the best result for the CT input with 0.96±0.04 validation AUC and 90%± 6%validation accuracy. The KNN model showed the best result for the PET input with 0.88±0.07 validation AUC and 85%±9% validation accuracy.
Conclusion: Radiomics features can differentiate between FDG-avid breast cancer metastatic axillary lymphadenopathy and FDG-avid axillary lymphadenopathy that is reactive to the COVID-19 mRNA vaccine. Such a model may have a role as a decision support tool for radiologists and nuclear medicine physicians in this diagnostic dilemma.