nnU-Net-driven lymph node segmentation and pyradiomics-based feature extraction for automated extranodal extension classification in head and neck cancer

Extranodal extension (ENE) is a critical prognostic factor in head and neck cancer. This study investigates the application of deep learning-based autosegmentation and radiomic analysis for ENE classification, aiming to enhance diagnostic accuracy through a two-stage approach combining advanced image segmentation techniques with machine learning classifiers to create a fully automated classification pipeline. Our study addresses two key challenges in extranodal extension (ENE) classification: accurate lymph node segmentation and robust feature extraction for ENE prediction. We developed a two-stage approach combining deep learning-based segmentation with radiomic analysis. First, we trained an nnU-Net auto segmentation algorithm on a diverse dataset of 621 head and neck cancer patients with large suspicious cervical lymph nodes that were segmented by experienced head and neck radiologists, combining open-source data with clinical data from Cleveland Clinic and Emory University. After training, we then evaluated the segmentation model on 53 test patients with pathologically proven ENE status. Following automated segmentation of the test set, we extracted radiomic features from the automated segmentations using the open-source Pyradiomics software and performed five-fold cross-validation using four machine learning classifiers to characterize ENE positive (patients who have ENE) and ENE negative (patients who do not have ENE) patients: Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost. Least absolute shrinkage and selection operator regularization (LASSO) was employed for feature selection for the logistic regression model, identifying the 10 most relevant radiomic features. Receiver operating characteristic (ROC) curves were then generated, and area under the curve was calculated for each classifier. Feature importance was assessed based on the absolute values of the LASSO coefficients. The Youden index was used to determine optimal classification thresholds for sensitivity and specificity. A total of 25 ENE positive and 28 ENE negative patients were included in the test set. The auto segmentation of the largest cervical lymph node achieved a Dice score of 0.84 ± 0.09 and a Hausdorff Distance of 38.85 ± 57.92. In the classification stage, Logistic Regression of the top ten pyradiomic features demonstrated the best overall performance with an AUC of 0.84 ± 0.13, an accuracy of 0.85 ± 0.10, a sensitivity of 0.84 ± 0.15, and a specificity of 0.85 ± 0.15. Random Forest achieved a lower AUC of 0.66 ± 0.12, with an accuracy of 0.72 ± 0.06, a sensitivity of 0.68 ± 0.20, and a specificity of 0.76 ± 0.13. SVM performed relatively well with an AUC of 0.77 ± 0.08, an accuracy of 0.79 ± 0.07, a sensitivity of 0.68 ± 0.16, and a specificity of 0.89 ± 0.14. XGBoost showed the lowest performance with an AUC of 0.51 ± 0.23, an accuracy of 0.66 ± 0.09, a sensitivity of 0.56 ± 0.32, and a specificity of 0.73 ± 0.21. The most important radiomic features were predominantly wavelet-based and included wavelet based gray level size zone matrix as the best predictor, highlighting the importance of heterogeneity in attenuation within the lymph node. Combining fully automated segmentation and radiomic analysis demonstrates significant potential for enhancing ENE classification accuracy in head and neck cancer. The Logistic Regression model, in particular, shows promise with its balanced performance across various metrics. By leveraging advanced auto segmentation and radiomic feature extraction, our method could provide valuable decision support for clinicians in a fully automated manner, potentially improving diagnostic precision and prognostic assessment.