Segmentation Variability in Bayesian U-Net versus Manual Annotations: Impact on Radiomic Reproducibility in Lung...

Area of Science:

• Medical imaging
• Radiology
• Artificial intelligence in medicine

Background:

• Radiomic analysis relies heavily on accurate segmentation of Regions of Interest (ROIs).
• Deep learning (DL) offers automated segmentation, potentially improving radiomic reproducibility.
• Quantifying uncertainty in DL predictions can enhance trust, especially if it reflects expert variability.

Purpose of the Study:

• To assess if DL-derived uncertainty aligns with expert annotation variability.
• To determine optimal configurations for maximizing radiomic feature reproducibility using DL segmentation.
• To evaluate the Monte Carlo Dropout (MCD) approach for segmenting lung tumors on CT scans.

Main Methods:

• Integrated Monte Carlo Dropout (MCD) into a U-Net model for lung tumor segmentation on CT images.
• Utilized two public datasets with multiple expert-annotated tumor masks.
• Compared MCD-based segmentations at various confidence levels with expert delineations.
• Extracted radiomic features and assessed reproducibility across different confidence thresholds.

Main Results:

• MCD-based segmentation uncertainty partially reflects expert variability, particularly at lower confidence levels.
• Radiomic features remain sensitive to segmentation variability; only about half achieved reproducibility under optimal conditions.
• The study provides insights into the trustworthiness of DL predictions compared to manual delineation.

Conclusions:

• DL segmentation with uncertainty quantification, like MCD, can enhance the trustworthiness of radiomic analysis.
• While not perfectly replicating expert variability, MCD offers a promising approach for improving radiomic reproducibility.
• Further optimization is needed to fully leverage DL for consistent radiomic feature extraction in clinical practice.