Robustness of Radiomic Features: Two-Dimensional versus Three-Dimensional MRI-Based Feature Reproducibility in...

Area of Science:

• Radiology
• Medical Imaging
• Oncology

Background:

• Radiomic features derived from medical imaging hold potential for non-invasive tumor characterization.
• Assessing the reproducibility of these features is crucial for their clinical translation.
• Lipomatous soft-tissue tumors present a unique challenge for imaging-based analysis.

Purpose of the Study:

• To compare the intra- and inter-observer variability of radiomic features extracted from two-dimensional (2D) and three-dimensional (3D) magnetic resonance imaging (MRI) in lipomatous soft-tissue tumors.
• To evaluate the impact of segmentation method (2D vs. 3D) on feature reproducibility, robustness, and the diagnostic performance of machine learning models.

Main Methods:

• A retrospective analysis of MRI scans from patients with histopathologically confirmed lipomatous soft-tissue tumors.
• Manual tumor segmentation was performed by two observers on both 2D and 3D MRI datasets.
• Radiomic features were extracted, and their reproducibility was assessed using the intraclass correlation coefficient (ICC > 0.75).
• Machine learning models were trained using reproducible features to evaluate diagnostic accuracy, sensitivity, and specificity.

Main Results:

• Three-dimensional (3D) segmentation yielded a significantly higher rate of reproducible radiomic features (58.9%) compared to two-dimensional (2D) segmentation (37.6%).
• A greater proportion of 3D features exhibited moderate-to-high robustness, whereas 40.9% of 2D features had low robustness.
• Machine learning models showed comparable diagnostic accuracy (90% for 3D vs. 88% for 2D), but 3D segmentation achieved higher sensitivity (83% vs. 75%) with equal specificity (100%).

Conclusions:

• Both 2D and 3D MRI-based radiomic features are reproducible for lipomatous soft-tissue tumors.
• Three-dimensional (3D) contour-focused segmentation demonstrates superior performance in terms of stable feature rate and feature robustness.
• The findings support the selection of 3D segmentation for enhanced feature extraction and potentially improved diagnostic performance in radiomic studies of these tumors.