Investigation of thoracic four-dimensional CT-based dimension reduction technique for extracting the robust radiomic...

Area of Science:

• Radiomics
• Medical Imaging
• Computational Pathology

Background:

• Radiomic feature selection traditionally relies on RIDER test-retest datasets.
• Discrepancies in CT protocols between facilities and test-retest datasets pose challenges.
• Four-dimensional CT (4D-CT) scans offer an alternative for robust feature selection in radiation therapy patients.

Purpose of the Study:

• To investigate the feasibility of using thoracic 4D-CT scans for robust radiomic feature selection.
• To compare feature robustness derived from 4D-CT data with traditional test-retest datasets.
• To identify optimal breathing phases within 4D-CT for reliable feature extraction.

Main Methods:

• Extracted 1170 radiomic features (shape, statistics, texture) from 4D-CT datasets of 14 lung cancer patients and 14 test-retest datasets.
• Utilized concordance correlation coefficient (CCC) > 0.85 as the threshold for robust feature selection.
• Analyzed feature robustness across different breathing phases (10 phases vs. 40%-60% phases) in 4D-CT data.

Main Results:

• Robust features ranged from 72% to 83% in 4D-CT groups using 40%-60% breathing phases.
• In contrast, using 10 breathing phases yielded only 4% to 41% robust features in 4D-CT.
• The number of robust features from 4D-CT (using 40%-60% phases) closely matched test-retest results (83%).

Conclusions:

• Respiratory motion significantly impacts radiomic feature robustness in 4D-CT.
• Employing 40%-60% breathing phases in 4D-CT enables robust feature selection comparable to test-retest datasets.
• This approach prevents excessive dimensionality reduction and facilitates selection of facility-protocol-specific robust features.