Enhancing online adaptive radiotherapy with uncertainty based segmentation error and out-of-distribution detection.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Anatomical segmentation introduces significant uncertainty in online adaptive radiotherapy.
• Deep learning (DL) models are increasingly used for segmentation in radiotherapy workflows.
• Accurate segmentation is crucial for precise radiation delivery and patient safety.

Purpose of the Study:

• To investigate the relationship between estimated uncertainty from DL segmentation and segmentation accuracy.
• To evaluate the ability of uncertainty estimation to detect out-of-distribution (OOD) data.
• To assess the utility of uncertainty estimation for quality control in adaptive radiotherapy.

Main Methods:

• Applied Monte Carlo dropout to a DL model for segmenting prostate cancer images (clinical target volume, bladder, rectum) from MR-guided radiotherapy.
• Utilized predictive entropy (PE) to quantify model and data uncertainty, establishing a threshold to classify segmentations as 'certain' or 'uncertain'.
• Employed mutual information (MI) to differentiate in-distribution (ID) from OOD data using healthy volunteer MRI scans.

Main Results:

• The DL segmentation model achieved high Dice scores (e.g., 94.8% for bladder).
• Higher PE values correlated with segmentation borders and incorrect predictions, enabling detection of errors.
• Mutual information achieved 100% separation between ID and OOD data, demonstrating robustness against unexpected inputs.

Conclusions:

• Uncertainty estimation in DL segmentation for MR-guided prostate cancer radiotherapy correlates with segmentation accuracy (Dice scores).
• This approach shows promise for real-time quality assessment of segmentations within the adaptive radiotherapy workflow.
• Preliminary findings suggest uncertainty estimation can effectively differentiate between expected (ID) and unexpected (OOD) data.