Related Experiment Video
Updated: Jun 23, 2025
10:17
Guidelines and Experience Using Imaging Biomarker Explorer IBEX for Radiomics
Published on: January 8, 2018
13.2K
Radiomics as a measure superior to common similarity metrics for tumor segmentation performance evaluation
Rukhsora Akramova1, Yoichi Watanabe1
1Department of Radiation Oncology, University of Minnesota, Minneapolis, Minnesota, USA.
Journal of Applied Clinical Medical Physics
|June 26, 2024
Summary
Radiomics features, measured by intraclass correlation coefficient (ICC), are superior to Dice similarity coefficient (DSC) and Hausdorff distance (HD) for evaluating tumor segmentation accuracy. This approach offers more sensitive detection of subtle segmentation variations in lung cancer patients.
Area of Science:
- Medical imaging analysis
- Radiomics and quantitative imaging
- Computational pathology
Background:
- Accurate tumor segmentation is crucial for treatment planning and response assessment in lung cancer.
- Traditional metrics like Dice similarity coefficient (DSC), surface Dice similarity coefficient (sDSC), and Hausdorff distance (HD) have limitations in capturing subtle segmentation differences.
- Radiomics features offer a quantitative approach to characterizing tumor heterogeneity and shape.
Purpose of the Study:
- To propose radiomics features, evaluated using the intraclass correlation coefficient (ICC), as a superior metric for assessing segmentation ability.
- To compare the performance of radiomics features with commonly used metrics (DSC, sDSC, HD) in evaluating segmentation accuracy.
- To determine if radiomics features can detect subtle variations missed by conventional metrics.
Main Methods:
- Extracted radiomics features from 90 segmented lung tumors across 10 patients using PyRadiomics.
- Calculated ICC for radiomics features to assess segmentation similarity.
- Compared ICC values with DSC, sDSC, and HD calculated from nine segmentations per tumor.
Main Results:
- Radiomics features, assessed by ICC, demonstrated higher sensitivity to segmentation changes compared to DSC and sDSC.
- Specific radiomics features (wavelet-LLL) showed ICCs ranging from 0.033 to 0.998.
- Conventional metrics showed DSC > 0.778, sDSC > 0.700, and HD varied from 0 to 1.9 mm, indicating less sensitivity to subtle variations.
Conclusions:
- Radiomics features with ICC provide a more sensitive and comprehensive evaluation of tumor segmentation ability than traditional metrics.
- This approach is valuable for assessing physician segmentation skills and auto-segmentation tool performance.
- The proposed radiomics-based evaluation can enhance medical training and the development of new auto-segmentation methods.
