Assessing robustness of radiomic features by image perturbation

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Computational Pathology

Background:

• Reproducibility in radiomics requires image features robust to variations in positioning, acquisition, and segmentation.
• Non-robust features in radiomic models can lead to inaccurate outcome predictions.
• Test-retest imaging is the standard for assessing feature robustness but is not always feasible.

Purpose of the Study:

• To investigate image perturbation combinations as an alternative to test-retest imaging for evaluating radiomic feature robustness.
• To compare the robustness of features identified through perturbations versus test-retest imaging.

Main Methods:

• Evaluated 18 image perturbation combinations (noise, translation, rotation, volume change, contour randomization) on CT images from 31 non-small-cell lung cancer (NSCLC) and 19 head-and-neck squamous cell carcinoma (HNSCC) patients.
• Computed 4032 morphological, statistical, and texture features from gross tumor volumes.
• Assessed robustness using the 95% confidence interval (CI) of the intraclass correlation coefficient (ICC), with CI ≥ 0.90 considered robust.

Main Results:

• The NTCV, TCV, RNCV, and RCV perturbation chains identified the fewest false positive robust features (NSCLC: 0.2-0.9%; HNSCC: 1.7-1.9%).
• These perturbation methods yielded results comparable to test-retest imaging in assessing feature robustness.
• A high degree of feature robustness was observed across tested perturbations.

Conclusions:

• Specific image perturbation chains (NTCV, TCV, RNCV, RCV) effectively assess radiomic feature robustness.
• These perturbation methods serve as a practical alternative to test-retest imaging, especially when test-retest data is unavailable.
• Ensuring feature robustness is key for reliable radiomic model development and clinical application.