Program Information
The Impact of Deformable Image Registration On the Robustness of Radiomic Features
K Chou1 , K Latifi2 , E Moros3*, V Feygelman4 , T Huang5 , G Zhang6 , (1) China Medical University, Taichung, Taiwan, (2) H. Lee Moffitt Cancer Center, Tampa, Florida, (3) H. Lee Moffitt Cancer Center, Tampa, FL, (4) H. Lee Moffitt Cancer Center, Tampa, FL, (5) China Medical University, Taichung, Taiwan, (6) H. Lee Moffitt Cancer Center, Tampa, FL
Presentations
TU-H-FS4-6 (Tuesday, August 1, 2017) 4:30 PM - 6:00 PM Room: Four Seasons 4
Purpose: Radiomics has been increasingly used for outcome prediction, treatment monitoring, lesion classification or potential imaging-biomarker. Deformable image registration (DIR) is often used in radiotherapy to integrate multi-modality images (e.g. from MRI, PET, CT, etc)for multiple uses before, during and after the treatment process. This study investigated the impact of DIR on radiomics features after contour propagation and image deformation.
Methods: 23 lung cancer patients (25 lesions) underwent 4D-CT scan with the same protocol. For each patient, the contour of each lesion was delineated on CT50%phase (Contour50%phase) and then propagated onto CT0%phase to generate Contourmapped-0%phase. The CT0%phase image was mapped to CT50%phase to generate CTmapped-50%phase image. Radiomic features (shape, intensity, LoG, wavelets, Laws, Co-occurrence, Run-length, Gray-level size-zone, Neighborhood Gray-tone Difference (NGTD) and fractal dimension) were computed within each group, standard group (Contour50%phase / CT50%phase); Contour-mapped group (Contourmapped-0%phase / CT0%phase) and Image-mapped group (Contour50%phase / CTmapped-50%phase). The relative difference (Diff) and Spearman’s rank correlation coefficient (SCC) were calculated between each group and standard (Contour50%phase / CT50%phase). The features of which Diff was less than 10% and SCC above 0.75 were considered robust and the percentage of robust features (PRF) for each type of feature was calculated.
Results: The PRF of shape, intensity, wavelet, Co-occurrence, Run-length, NGTD and fractal dimension features in Image-mapped were more than 70%. However, in Contour-mapped group, the intensity and Co-occurrence features were the only two categories with PRF more than 70%, and no NGTD feature was robust. For the subtypes, LoG_0.5, Wavelet_HHH and Laws_LLL, were the only robust features that demonstrated robustness both in Contour-mapped and Image-mapped.
Conclusion: Propagated contour and mapped image using DIR may lead variations in feature extraction. But there are still a few types and subtypes of features that are robust in both ways.
Contact Email: