MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Early Prediction of Locoregional Recurrence in Head & Neck Cancer After Radiation Therapy Through Multifaceted Radiomics

Z Zhou1*, D Sher2 , Q Zhang3 , J Shah4 , N Pham5 , L Chen6 , M Folkert7 , S Jiang8 , J Wang9 , (1) UT Southwestern Medical Center, Dallas, Texas, (2) UT Southwestern Medical Center, Dallas, Texas, (3) UT Southwestern Medical Center, Dallas, Texas, (4) UT Southwestern Medical Center, Dallas, Texas, (5) UT Southwestern Medical Center, Dallas, Texas, (6) UT Southwestern Medical Center, Dallas, Texas, (7) UT Southwestern Medical Center, Dallas, Texas, (8) UT Southwestern Medical Center, Dallas, TX, (9) UT Southwestern Medical Center, Dallas, TX

Presentations

(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM

Room: Room 207

Purpose: PET and CT are routinely used for post-radiotherapy treatment follow-up in head & neck (H&N) cancer. We developed a novel unified framework termed as multifaceted radiomics (M-radiomics) for locoregional recurrence (LR) prediction through analyzing quantitative imaging features extracted from the first follow-up PET and CT after definitive radiation therapy.

Methods: FDG-PET and CT from 100 H&N patients with definitive radiation therapy were retrospectively selected for this study. Among these patients, 40 experienced LR. Two hundred and fifty-seven features (including intensity, texture and geometry) were extracted from contoured tumors on PET and CT, respectively. These features were then used as the initial input for the M-radiomics model for predicting LR. The M-radiomics optimally combines the output from multiple modality-specific radiomic models. In both CT- and PET-based models, multiple criteria including sensitivity and specificity are considered simultaneously during the model training. Additionally, different types of classifiers, including support vector machine (SVM), logistic regression (LR), discriminant analysis (DA), decision tree (DT), K-nearest-neighbor (KNN), and naive Bayesian (NB), are used for each imaging modality. M-radiomics considers multi-modality, multi-classifier and multi-criteria into a unified framework to obtain more reliable prediction results by fusing outputs from different modality-specific classifiers. Specifically, a reliable classifier fusion (RCF) strategy is developed, which not only considers the relative importance among different classifiers, but also considers the reliability of the classifier itself.

Results: The accuracies for PET, CT, PET & CT are 73.00%, 75.00%, 78.00%, and the AUC are 0.7473, 0.7633, 0.7848, respectively. The sensitivities for three modalities are 65.00%, 65.00%, 65.00%, while the specificities are 78.33%, 81.67% and 86.67%.

Conclusion: A new M-radiomics framework which considers multi-modality, multi-classifier and multi-criteria into a unified model was proposed. Compared with radiomic model based on single modality or classifier, the M-radiomics can obtain more accurate and reliable predictive results.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email