Machine learning methods for optimal differentiation of recurrence versus inflammation from post-therapy nasopharyngeal ¹⁸F-FDG PET/CT images

Objectives: Despite significant improvements in local control due to improved radiotherapy, local recurrence remains a primary issue in patients with advanced nasopharyngeal carcinoma (NPC). The purpose of this study was to identify optimal machine learning methods for radiomics-based differentiation of local recurrence versus inflammation from post-treatment nasopharyngeal PET/CT images.

Methods: 76 patients treated for NPC were enrolled in our study (41/35 local recurrence/ inflammation as confirmed by pathology). A total of 80 radiomic features were extracted from ¹⁸F-FDG PET images for each patient. The diagnostic performances (differentiating between local recurrence vs. inflammation) were investigated for 84 cross-combinations derived from 12 feature selection methods and 7 classifiers. Leave-one-out cross validation was applied for feature selection and classification, and the diagnostic performance was evaluated by two metrics: area under the receiver operating characteristic (ROC) curve (AUC) and test error. Furthermore, we compared the performance of radiomic signature with routine SUVmax.

Results: (1) 5/7 classifiers (KNN, LDA, LR, NB, RBF-SVM) exhibited high diagnostic performance in combination with the majority of feature selection methods. (2) Most combined machine learning methods demonstrated relatively low test error except for 6/12 combinations with the DT classifier and LS + KNN. (3a) 10 combined methods satisfied both AUC>0.8592 (75th percentile value) and test error<0.2237 (25th percentile value), but only two combinations also met the sum of sensitivity and specificity>1.6599 (75th percentile value). (3b) FSV + KNN and L0 + NB displayed performances that were overall more accurate (AUC, 0.8857 and 0.8878; sensitivity, 0.7317 and 0.7805; specificity, 0.9429 and 0.9143) and reliable (test error, 0.2105 and 0.2105) relative to other methods. (4) The radiomic signature performed significantly better than SUVmax alone (AUC: 0.8369; sensitivity: 0.7317; specificity: 0.8857; test error: 0.2763), the p-value of AUC less than 0.05.

Conclusions: Our study identified the most accurate and reliable machine learning methods for radiomics-based differentiation of local recurrence versus inflammation from post-treatment nasopharyngeal PET/CT images. The resulting radiomic signature significantly outperformed conventional analysis in terms of accuracy and reliability. Acknowledgments: This work was supported by the National Natural Science Foundation of China under grants 61628105, 81501541, the National key research and development program under grant 2016YFC0104003, the Natural Science Foundation of Guangdong Province under grants 2016A030313577, and the Program of Pearl River Young Talents of Science and Technology in Guangzhou under grant 201610010011.