Segmentation Uncertainty Estimation as a Sanity Check for Image Biomarker Studies

Area of Science:

• Radiomics
• Medical Image Analysis
• Cancer Prognosis

Background:

• Radiomics, or image biomarker analysis, aids in tissue characterization and prognosis using clinical images.
• Reproducibility issues in radiomics stem from uncertainties in image acquisition, processing, and segmentation.
• Harmonization techniques aim to reduce these uncertainties and improve prognostic model performance.

Purpose of the Study:

• To estimate protocol-induced uncertainty in image biomarkers.
• To assess the impact of segmentation uncertainty on prognostic model performance.
• To optimize radiomics models by accounting for segmentation variability.

Main Methods:

• Utilized two non-small cell lung cancer (NSCLC) cohorts (421 and 240 patients) for training and testing.
• Employed a Monte Carlo algorithm to generate 300 synthetic contours per patient, assessing sensitivity (η) to contour perturbation.
• Developed Cox proportional hazards models using low-η (stable) and high-η (unstable) radiomic features, tested across 5000 augmented realizations.

Main Results:

• Low-η features yielded stable prognostic models, achieving p < 0.05 in 90% of augmented realizations.
• High-η features resulted in unstable models, with p < 0.05 in only 30% of realizations.
• Low-η models showed prediction uncertainty in 10% of cases, compared to 50% for high-η models.

Conclusions:

• Relying solely on original segmentations for radiomics model evaluation can be misleading due to inherent segmentation uncertainty.
• Simulating segmentation uncertainty is crucial for developing stable and reliable image biomarker selection and prognostic models.
• The proposed method for estimating segmentation uncertainty is universal and adaptable for other protocol uncertainties.