Locoregional Recurrence Prognostication in Head and Neck Cancer patients: A multi-modality PET/CT Fusion Radiomics Study

Introduction: Locoregional Recurrence (LRR) plays an inauspicious role in head and neck cancer (HNC) treatment failure. The development of prognostic models for the prediction of LRR in HNC patients is highly desired. Noninvasive/inexpensive characterization of inter/intra tumor heterogeneities by quantitative analysis of medical images (radiomics) have provided new opportunities toward personalized treatment planing. In this study, we aimed to find optimum radiomics prognostic model for identification of HNC patients at high risk of LRR, using single- and multimodality PET/CT fusion, along with multiple feature selection (FS) and machine learning (ML) methods.

Methods: The dataset comprised 201 HNC patients (21 of which experienced LRR after treatment) from four different centers, which underwent PET/CT scan prior to treatment. Tumor segmentation was performed on PET/CT scans by an experienced nuclear medicine physician. Images were registered and interpolated to isotropic voxel spacing of 1mm³ prior to fusion. Beside single CT and PET scans, two multimodality scans, one fused using Non-Subsampled Contourlet Transform (NSCT) and the other using visual saliency map and Weighted Least Square (WLS) optimization were analyzed. Radiomics features were extracted via Standardized Environment for Radiomics Analysis (SERA) package which agrees with Image Biomarker Standardized Initiative (IBSI) guideline. Prognostic models were developed by cross combinations of five FS and five ML methods, all capable of continuous time-to-event prediction modeling. FS methods included Mutual Information (MI), Variable Hunting (VH), Variable Hunting with Variable Importance (VH.VIMP), tree minimal depth methodology (MD), and selecting top ten features with highest univariate performance via Cox regression (Cindex). ML methods were Cox Proportional Hazard (CoxPH), Cox model fitted by likelihood-based boosting (CoxBoost), Lasso and Elastic-Net regularized generalized linear model (glmnet), random survival forest (RSF), and gradient boosting with component-wise linear model (glmboost). Models were trained and their hyperparameters were optimized (using grid search) using 70% of the dataset which was randomly selected. The trained model was applied to the rest of datasets (unseen by the model during training), with 1000 bootstraps and the average performance of the model was reported to avoid unwanted biases. Performance of the models were reported with Concordance Index (C-index).

Results: Regarding the comparison of the different scans, multimodality NSCT fused images achieved the highest average performance (average C-index: 0.75). In single-modality models, CT images were best trained with MD_glmnet (as FS_ML method) and achieved C-index of 0.76, while PET scans revealed their best performance via RSF_VH.VIMP (C-index: 0.78). Among multimodality models, WLS scan trained with Cindex_CoxPH (C-index: 0.84) and NSCT scan trained with VH-CoxPH (C-index: 0.82) achieved highest performances.

Conclusions: Compared to radiomics signatures based on single PET or CT scans, feature signatures extracted from multimodality fused scans revealed superior performance for prediction of locoregional recurrence in HNC patients. Different modalities achieved their highest performance with different combination of FS and ML methods, meaning that there is no fit-to-all algorithm. Models developed in this study can be used by clinicians as a decision-support system to establish personalized treatment regimens.

• Download figure
• Open in new tab
• Download powerpoint