Click here to


Are you sure ?

Yes, do it No, cancel

Multifaceted Radiomics: Towards More Reliable Radiomics for Predicting Distant Metastasis in Head & Neck Cancer

Z Zhou*, K Wang , H Liu , D Sher , J Wang , UT Southwestern Medical Center, Dallas, TX


(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: Reliably predicting distant metastasis (DM) is critical for stratifying high-risk head & neck (H&N) patients to receive intensified systemic therapy with potentially improved survival. We aim to develop a novel multifaceted radiomics (M-radiomics) model for DM prediction using PET and CT images.

Methods: Totally 188 patients with FDG-PET, radiotherapy planning CT and clinical parameters from TCIA were used in this study. Before training and testing, radiomic features (including intensity, texture and geometry) are extracted from PET and CT images, and deep autoencoder is adopted to transform features from three modalities (clinical parameter, PET and CT features) into one representation feature set. The training stage consists of classifier specific model construction, and Pareto-optimal model set generation. To increase the diversity, three classifiers including support machine, random forest and Adaboosting are adopted to build predictive models using the representation features. An iterative multi-objective immune algorithm is employed to optimize the multi-objective model that considers sensitivity and specificity as the objective functions simultaneously. Then three classifier-specific Pareto-optimal model sets are generated and the final Pareto-optimal model set is generated through sorting the models in the aforementioned three Pareto-optimal model sets in a non-domainted way. The testing stage consists of weight calculation and evidential reasoning (ER) based fusion. After setting the zero weights for the models with extremely imbalanced sensitivity and specificity, the weights for the remaining models are calculated based on the area under the curve (AUC). The final predicted probabilities are obtained by fusing the probabilities of all the Pareto-optimal models through ER approach.

Results: AUC, accuracy, sensitivity, and specificity for M-radiomics are 0.845, 0.809, 0.800, and 0.810, respectively. It outperforms multi-objective radiomics model achieving 0.750 AUC.

Conclusion: We developed a reliable and automatic M-radiomics for predicting DM in H&N cancer. The results demonstrated that M-radiomics outperformed the traditional radiomics model.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the US National Institutes of Health (R01 EB020366)


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email