ARRS 2022 Abstracts


1954. Impact Of Simulated Marginal Erosions of Volumetric Segmentation of Pancreatic Adenocarcinoma (PDA) on the Robustness of Radiomics Features
Authors * Denotes Presenting Author
  1. Leonardo Iheme; Mayo Clinic
  2. Anurima Patra; Mayo Clinic
  3. Garima Suman; Mayo Clinic
  4. Hala Khasawneh; Mayo Clinic
  5. Sovanlal Mukherjee *; Mayo Clinic
  6. Panagiotis Korfiatis; Mayo Clinic
  7. Ajit Goenka; Mayo Clinic
Radiomics is increasingly being used for outcomes prediction in PDA. The infiltrative morphology of PDA results in variations in the margins of tumor segmentation, which has potential to impact the robustness and task-specific performance of features extracted from such segmentations. Our purpose was to evaluate the impact of variations in the margins of segmented PDA on the robustness of radiomics features.

Materials and Methods:
A Medical Imaging Data Readiness Scale (MIDaR) level A dataset (portal venous phase CTs, slice thickness = 3.75 mm) of 518 treatment-naïve biopsy-proven PDA was created after exclusion of CTs with suboptimal image quality or biliary stents. Volumetric gross tumor volumes (GTV) on CTs done by two radiologists using 3D Slicer and subsequently reviewed by a senior radiologist were the ground truth (GT). Images were pre-processed with a modified soft tissue CT window (level 100 HU, width 400 HU) and gray level discretization into bins of width 25. To simulate variations in segmentation margins, marginal erosions (5% increments, range 5-20, E5-E20) were applied to volumetric GT segmentations. Using PyRadiomics, 18 first order and 70 gray level (GL) features were extracted. Features from each eroded mask were compared with features extracted from the corresponding GT. Robustness was quantified as the proportion of features that fell into distinct ranges of the concordance correlation coefficient (CCC): excellent (CCC>0.85), good (0.7<=CCC<=0.85), moderate (0.5<=CCC<0.7), and poor (CCC<?0.5).

Mean (range) tumor diameter on GT was 6.1 cm (1.8, 16.5) cm whereas it was 6.1 cm (1.8, 16.3) cm, 5.8 cm (1.8, 16.2) cm, 5.9 cm (1.8, 16.1) cm, and 5.8 cm (1.8, 16.0) cm for E5-E20, respectively. Dice scores [mean (standard deviation)] for E5-E20 were 0.97 (0.00), 0.95 (0.00), 0.92 (0.00) and 0.89 (0.00), respectively. As a group, both first order (CCC: 0.99, 0.98, 0.96 and 0.94 for E5-E20, respectively) and GL features (CCC 0.98, 0.94, 0.91 and 0.88 for E5-E20, respectively) had excellent robustness. Similar trend was noted for individual groups of GL features: GL Co-occurrence Matrix (GLCM) (CCC from 0.99 to 0.92), GL Run Length Matrix (CCC from 0.97 to 0.86), GL Size Zone Matrix (CCC from 0.96 to 0.85) and GL Dependence Matrix (CCC from 0.96 to 0.85).

Based on a study of simulated volumetric segmentations in a large, curated dataset, we conclude that up to 20% erosion at the edges has minimal impact on the robustness of first order and GL radiomics features extracted from PDA. Since marginal erosions simulate exclusion of tumor penumbra during PDA segmentation, accurate segmentation of tumor core is likely more critical for generation of reproducible radiomic models. Further investigation is needed to assess the impact of these findings on task-specific performance of radiomics features.