CLEAR item#19

“Segmentation strategy. Indicate which software programs or tools are used for segmentation or annotation. Specify the version of the software and the exact configuration parameters. Provide reference and web link to the software. Describe the segmentation method (e.g., automatic, semi-automatic, manual). Provide the rules of the segmentation (e.g., margin shrinkage or expansion from the visible contour, included/excluded regions). Provide figures to show the segmentation style. Provide image registration details (e.g., software, version, link, parameters) if segmentation is propagated for multi-modal (e.g., CT and MR), multi-phase (e.g., unenhanced, arterial, venous phase CT), or multi-sequence (e.g., T2-weighted, post-contrast T1-weighted, diffusion-weighted imaging) analyses. If radiomic features are extracted from 2D images on a single slice, please explain with which criteria the slice is chosen. In the case of several lesions, explain if all the lesions are segmented and describe how the feature values are aggregated. If only one lesion is chosen, describe the criteria (e.g., the primitive or the most voluminous).” [1] (from the article by Kocak et al.; licensed under CC BY 4.0)

Reporting examples for CLEAR item#19

Example#1. “In the discovery cohort, the CT scan images were pseudonymised and transferred to the eXtensible Neuroimaging Archive Toolkit (XNAT) platform for image curation and segmentation. Imaging data were in the Digital Imaging and Communications in Medicine format. In the validation cohort, pseudonymised CT scan images from the European Organisation for Research and Treatment of Cancer repository were uploaded onto XNAT. CT scan variables are listed in the appendix (pp 7–8). The tumour was manually delineated on all slices using XNAT contouring tools producing 3D regions of interest. Three independent users completed whole lesion segmentations (experienced sarcoma radiologist CK-M, clinical fellow AA, and senior research radiographer RE). Segmentations completed by RE were reviewed by AA and an independent senior sarcoma radiologist (CM). A semiautomated sub-segmentation tool was used to obtain radiomic volume fraction (RVF) features for four sub-regions associated with low, middle, high, and very high Hounsfield units. Sub-segmentations were obtained semi-automatically using the algorithm described in the appendix (p 1), which makes use of patient-specific Hounsfield unit thresholds in conjunction with morphological operations with the aim of generating sub-segmentations that are similar to those a human user would create manually. Because of challenges in automatically identifying the visually apparent high Hounsfield unit regions in some tumours, guide regions of interest were drawn on a single slice of each contiguous high Hounsfield unit sub-region. An alternative approach for computing the sub-region volume fractions was developed to reduce user subjectivity. The RVF feature values obtained from the algorithm were used to derive fixed Hounsfield unit thresholds, from which approximate RVF (ARVF) estimates were computed by simple image thresholding. This procedure is detailed in the appendix (pp 2–3, 14–15) and applied to the discovery dataset only to ensure that the estimates of predictive performance obtained using the validation dataset are not biased to user subjectivity. The thresholds thus derived were –50, 19, and 228 Hounsfield units (low is <–50, middle is –50 to <19, high is 19 to <228, very high is ≥228).” [2] (from the article by Artur et al.; licensed under CC BY-NC-ND 4.0)

Example#2. “Tumor segmentation: Radiologist 1 performed three-dimensional volume of interest (3D-VOI) segmentation along the tumor margin excluding cysts, necrosis, blood vessels, and lymph nodes inside tumor on axial portal venous phase CT images using ITK-SNAP software (version 3.8.0, http://www.itksnap.org/). We did not choose the arterial phase because the tumor boundary was more distinct and evident in the portal venous phase, which contribute to tumor segmentation.” [3] (from the article by Liao et al.; licensed under CC BY 4.0)

Example#3. “The contrast-enhanced CT images at the arterial phase were uploaded to ITK-SNAP (version 3.6.0) software for tumor segmentation. A radiologist with 3 years of experience, who was blinded to the pathological results, manually delineated the maximum five slices of the tumor area. The segmentation was subsequently verified by another radiologist with 14 years of experience. We segmented the area of interest to exclude vascular structures and organs that were encased, resulting in only the tumor region being included for radiomics feature extraction. An example of tumor segmentation can be seen in Fig. 2. To ensure generalization performance, voxel resampling (1 × 1 × 1 mm3) was conducted on CT images prior to radiomics features extraction.” [4] (from the article by Wang et al.; licensed under CC BY 4.0)

Explanation and elaboration of CLEAR item#19

The significance of reporting the segmentation strategy is emphasized in Item#19. An extensive and complete explanation of the segmentation strategy is needed for multiple reasons. First, subtle differences in segmentation strategy significantly affect radiomic features and their predictive power [5]. Therefore, the effect of differences in segmentation strategy should be taken into account when designing and evaluating radiomics-based research methods. Second, data extracted from the segmented slice or volumes may vary depending on the type of segmentation (automated or not), technique and on the criteria used for assessing lesion margins, and this is challenging because many lesions show unclear borders. Therefore, readers need to know these fundamental data for reproducibility and for clinical transferability. Example#1 provides an extensive explanation of how the item should be completed. In addition, two additional concise examples are provided.

References

  1. Kocak B, Baessler B, Bakas S, et al (2023) CheckList for EvaluAtion of Radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights Imaging 14:75. https://doi.org/10.1186/s13244-023-01415-8
  2. Arthur A, Orton MR, Emsley R, et al (2023) A CT-based radiomics classification model for the prediction of histological type and tumour grade in retroperitoneal sarcoma (RADSARC-R): a retrospective multicohort analysis. Lancet Oncol 24:1277–1286. https://doi.org/10.1016/S1470-2045(23)00462-X
  3. Liao H, Yuan J, Liu C, et al (2023) Feasibility and effectiveness of automatic deep learning network and radiomics models for differentiating tumor stroma ratio in pancreatic ductal adenocarcinoma. Insights Imaging 14:223. https://doi.org/10.1186/s13244-023-01553-z
  4. Wang H, Xie M, Chen X, et al (2023) Radiomics analysis of contrast-enhanced computed tomography in predicting the International Neuroblastoma Pathology Classification in neuroblastoma. Insights Imaging 14:106. https://doi.org/10.1186/s13244-023-01418-5
  5. Poirot MG, Caan MWA, Ruhe HG, et al (2022) Robustness of radiomics to variations in segmentation methods in multimodal brain MRI. Sci Rep 12:16712. https://doi.org/10.1038/s41598-022-20703-9

Back Next