MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Predicting Acute-Phase Weight Loss Based On CT Radiomics and Dosiomics in Lung Cancer Patients Treated with Radiotherapy

S Lee , P Han , R Hales , K Voong , T McNutt , J Lee*, Johns Hopkins University, Baltimore, MD

Presentations

(Monday, 7/15/2019) 1:15 PM - 1:45 PM

Room: Exhibit Hall | Forum 2

Purpose: To analyze CT radiomics and dosiomics (R&D) features to predict acute-phase weight loss in lung cancer patients treated with radiotherapy.

Methods: Baseline weight of 319 lung cancer patients who underwent intensity modulated radiation therapy (IMRT) was measured between 1 month prior to and 1 week after the first IMRT. Their weight change between 1 week and 2 months after the first IMRT was analyzed, and dichotomized at 5% weight loss. Each patient had a planning CT and contours of gross tumor volume (GTV) and esophagus. A total of 350 features including 9 clinical, 62 GTV and esophagus (GTV&ESO) dose-volume histogram (DVH), 93 GTV radiomics, and 186 GTV&ESO dosiomics features were extracted. The R&D texture features were categorized as first- (L1), second- (L2) and higher-order (L3) statistics, and also 2 multilevel groups, L1+L2 and L1+L2+L3. Multilevel texture analysis was performed to identify optimal R&D input features by using nested cross-validation (NCV) with an outer 10×5-fold CV loop. NCV performance was compared using the area under the receiver operating characteristic curve (AUC) among 4 different input conditions: DVH-only, R&D-only, R&D+DVH, R&D+DVH+clinical features. Feature selection was performed with a series of all-relevant and minimal-optimal feature selection processes using Boruta algorithm followed by collinearity removal based on variance inflation factor and recursive feature elimination with adaptive LASSO-logistic regression. A 5-fold cross-validated glmnet model was used to assess predictive performance on outer CV test sets.

Results: Combined GTV&ESO L3 dosiomics and GTV L1+L2+L3 radiomics features were identified as optimal R&D input features. These optimal R&D input features achieved the best NCV performance with the training AUC of 0.73±0.02 (mean±SD) and test AUC of 0.71±0.06.

Conclusion: Using optimal R&D input features is beneficial for predicting early weight loss in lung cancer radiotherapy, and leads to improved prediction performance compared to using conventional DVH parameters.

Funding Support, Disclosures, and Conflict of Interest: This study was supported by Canon Medical Systems Corp.

Keywords

Image Analysis, Lung, Radiation Therapy

Taxonomy

TH- response assessment : Machine learning

Contact Email