Does consensus contour improve robustness and accuracy in ¹⁸F-FDG PET radiomic features?

Area of Science:

• Radiomics
• Medical Imaging
• Oncology

Background:

• Radiomic features extracted from medical images offer valuable quantitative information for disease characterization and treatment response assessment.
• The accuracy and robustness of these features are crucial for reliable clinical application, particularly in positron emission tomography (PET) imaging.
• Segmentation methods and initial mask choices significantly influence radiomic feature extraction.

Purpose of the Study:

• To validate the robustness and accuracy of consensus contour (ConSeg) for 2-deoxy-2-[$$F]fluoro-D-glucose ( $$F-FDG) PET radiomic features.
• To assess the impact of different segmentation methods and initial masks on radiomic feature reliability.
• To evaluate the performance of ConSeg compared to traditional segmentation approaches.

Main Methods:

• Utilized 225 nasopharyngeal carcinoma (NPC) and 13 extended cardio-torso (XCAT) simulated datasets.
• Applied four segmentation methods with two distinct initial masks for each dataset.
• Extracted 107 radiomic features using Pyradiomics and calculated intraclass correlation coefficients (ICC) to evaluate robustness and accuracy.

Main Results:

• Consensus contour (ConSeg) demonstrated higher ICC for radiomic features in robustness tests and comparable ICC in accuracy tests versus the average of individual segmentations.
• Irregular initial masks generally yielded higher ICC for both robustness and accuracy compared to rectangular masks.
• 19 features (17.76%) achieved an ICC ≥ 0.75 in both robustness and accuracy tests across various segmentation parameters.
• Dataset type exerted a substantial influence on the correlation relationships among radiomic features.

Conclusions:

• Combining consensus contour with irregular initial masks can partially enhance the robustness and accuracy of radiomic analysis.
• The correlation patterns of radiomic features are primarily dictated by the dataset, rather than the segmentation method or initial mask.
• These findings support the potential of ConSeg in improving the reliability of $$F-FDG PET-based radiomics.