Room: AAPM ePoster Library
Purpose: predict distant metastasis using Radiomics predictors extracted from relevant tumor subregions of CECT, CET1, and T2 image. To investigate the discriminability defined by the association of prognosis across different image modalities and different ROIs.
Methods: In total, 77 patients who underwent pre-treatment CECT, CET1, and T2 scans were enrolled retrospectively. Each tumor was partitioned manually into two subregions based on CET1 image. One is contrast-enhanced volume (CEV) and the other was non-contrast-enhanced volume (NCEV). There are three ROIs per patient including GTV and the two subregions. For each ROI, 107 features were extracted and Pearson Correlation was used to remove redundant features. LASSO was used for feature selection and model construction. Model-ensemble based iterative feature selection was used to remove feature iteratively and group of models were constructed across different feature set generated in every iteration. Average AUCs were used for evaluating the discriminability of different image modalities and different ROIs.
Results: of NCEV were located at the outer zone of tumor while CEV were located at the inner part. The NCEV (Testing AUC: 0.794, 0.824, 0.910, 0.943, 0.948 and 0.904) shows higher discriminability than CEV (Testing AUC: 0.696, 0.794, 0.837, 0.861, 0.916, and 0.850) and GTV (Testing AUC: 0.796, 0.704, 0.864, 0.830, 0.930, and 0.807) in CECT, CET1 and other modality combinations. The NCEV of CECT&CET1&T2 has the highest discriminability.
Conclusion: variations within the tumor possess higher prognostic potential than the whole tumor. This study may have significant implications for clinical oncology by identifying important tumor regions for biopsy. Furthermore, this is particularly relevant for radiotherapy treatment planning and adaptation, because high-risk tumor subregions associated with the aggressive disease can then be targeted with a radiation boost to potentially improve patient survival.