Stability of Multi-Parametric Prostate MRI Radiomic Features to Variations in Segmentation

Area of Science:

• Radiomics and Medical Imaging
• Computational Pathology
• Oncological Biomarker Development

Background:

• Radiomic feature stability is crucial for developing reliable imaging biomarkers in oncology.
• Segmentation variability among human annotators can compromise the reproducibility of radiomic features.
• Personalized oncological strategies rely on robust and stable imaging biomarkers.

Purpose of the Study:

• To develop an in silico method for simulating segmentation variations to assess radiomic feature stability.
• To identify stable radiomic features that are robust to segmentation differences.
• To evaluate the impact of pre-filtering strategies on radiomic feature stability.

Main Methods:

• Generated 15 synthetic contours by perturbing ground-truth prostate gland segmentations on multi-parametric MRI (T2w, ADC, SUB-DCE).
• Extracted 1224 radiomic features using Pyradiomics and assessed stability with ICC(1,1).
• Compared stable features across internal and external populations and investigated filtering strategies.

Main Results:

• Segmentation variations significantly impacted radiomic feature stability.
• Pre-filtering strategies substantially improved feature robustness: T2w (81% vs 36%), ADC (94% vs 36%), SUB-DCE (93% vs 43%).
• Identified robust features that remained stable across internal and external datasets.

Conclusions:

• In silico simulation of segmentation variations is effective for evaluating radiomic feature stability.
• Pre-filtering strategies are essential for mitigating the impact of segmentation variability on radiomic features.
• Robust radiomic features identified through this approach can enhance the reliability of imaging biomarkers for personalized cancer treatment.