Program Information
Quantifying Tumor Morphological Change with Jacobian Map for Prediction of Pathologic Tumor Response to Chemo-Radiotherapy in Locally Advanced Esophageal Cancer
S Riyahi Alam*, W Choi , C Liu , W Lu , Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
MO-F-CAMPUS-JT-5 (Monday, July 31, 2017) 4:30 PM - 5:30 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To quantify tumor morphological change due to chemo-radiotherapy (CRT) using Jacobian map computed from Deformation Vector Field (DVF). To extract quantitative radiomics features from the Jacobian map for the prediction of pathologic tumor response to CRT.
Methods: In 20 patients who underwent CRT, a free-form deformable registration with bending energy of transformation as regularizer was performed to register the follow-up (post-CRT) CT to the baseline CT image. Smoothness of transformation and improvement of alignment were evaluated using inverse consistency error and Normalized Mutual Information (NMI). Determinant of the gradient of the DVF was computed as Jacobian (J), which measured local volumetric shrinkage (J < 1) or expansion (J > 1). Radiomics features (intensity, texture, geometry) were then extracted from the Jacobian map to quantify tumor morphological change due to CRT. The importance of each feature in predicting pathologic tumor response was evaluated using the Least Absolute Shrinkage and Selection Operator (LASSO) and the Support Vector Machine (SVM) model with ten times 10 fold cross-validation.
Results: Inverse consistency error was 3.64±2.68 mm and average NMI after registration was improved from 0.15±0.02 to 0.25±0.03. Qualitatively, Jacobian map detected local shrinking/expanding regions in a tumor. Quantitatively, the average tumor volume ratio (i.e. mean J in the tumor) was 0.83±0.11 (17% shrinkage) and 1.11±0.16 (11% expansion) for responder and non-responder tumors (p = 0.0003), respectively. This agreed with the observation that responder tumors tended to have greater volumetric shrinkage. In uni-variate analysis, the two most significant predictors were Median J (p=0.0002, AUC=0.94) and Minimum J (p=0.002, AUC=0.89). In multivariate analysis, the best LASSO-SVM model used only two features (Median J and Minimum J) with high accuracy (Sensitivity=94.4%, Specificity=90.9%, AUC=0.91).
Conclusion: Novel radiomics features extracted from the Jacobian map quantified tumor morphological change and achieved high accuracy in predicting pathologic tumor response.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the National Cancer Institute Grants R01CA172638.
Contact Email: