Click here to


Are you sure ?

Yes, do it No, cancel

Multilevel Radiomics Analysis for Prediction of Baseline Pulmonary Function in Lung Cancer Patients Treated with IMRT

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


(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Room 205

Purpose: To identify an optimal combination of radiomic feature levels for predicting baseline pulmonary function in lung cancer patients treated with IMRT.

Methods: Forced expiratory volume in first second (FEV1) and forced vital capacity (FVC) were measured in 64 patients at baseline (between 3 months before and 1 week after their first IMRT). Each patient had a RT-planning CT and associated contours of gross tumor volume, left and right lungs, and esophagus. A total of 3,052 features were extracted from the original and wavelet-filtered CT images, and subdivided into five levels: shape (S), first-order (L1), second-order (L2: gray-level co-occurrence matrix features) and higher-order local texture features (L3: gray-level run-length and gray-level size-zone matrix features), and global texture features (G: 3D geometric moment invariants). Feature selection was based on a LASSO-logistic regression model embedded in a leave-one-out cross-validation (LOOCV) loop, following a series of univariate filtering with Spearman correlation and minimum-redundancy-maximum-relevance criterion in each fold. For model building, optimized signatures for predicting a FEV1/FVC ratio dichotomized at 70% were determined by sequential forward search of ranked features according to selected frequency in the LOOCV loop, among unilevel features or multilevel features extracted from each of four different groups: S+L1, S+L1+L2, S+L1+L2+L3 and S+L1+L2+L3+G. An optimized LASSO model was used to assess LOOCV-based performance.

Results: In unilevel analysis, the optimized L1 signature showed the best performance (AUC=0.901), followed by the optimized L2 (AUC=0.824), S (AUC=0.783), L3 (AUC=0.766), and G (AUC=0.661) signatures. In multilevel analysis, the optimized S+L1+L2+L3+G signature showed the best performance (AUC=0.944), followed by the optimized S+L1 (AUC=0.939), S+L1+L2 (AUC=0.926) and S+L1+L2+L3 (AUC=0.912) signatures.

Conclusion: The multilevel analysis strategy leads to superior performance, as compared to the unilevel approach. Global texture features should be combined with shape and local texture features to achieve the best performance.

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


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email