Improving confidence in MRI-based auto-segmentation via uncertainty assessment

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate delineation of organs of interest (OOIs) is critical for safe radiotherapy planning.
• Deep learning models achieve high geometric accuracy but often lack reliable uncertainty quantification, hindering clinical translation.
• Overconfident predictions in ambiguous regions necessitate improved model calibration for clinical workflows.

Purpose of the Study:

• To compare the standard nnU-Netv2 against its residual-encoding variant (ResEncM) for automated brain OOI segmentation.
• To evaluate if ResEncM offers superior reliability and calibration while maintaining geometric accuracy.
• To assess the clinical utility of uncertainty quantification in deep learning segmentation for radiotherapy.

Main Methods:

• T1-weighted contrast-enhanced MRI scans from 70 brain cancer patients were utilized.
• Ground-truth contours for brainstem, hippocampi, chiasm, optic nerves, optic tracts, and pituitary were delineated.
• Epistemic uncertainty was quantified using mutual information, and Expected Calibration Error (ECE) was computed.

Main Results:

• Both models achieved high geometric accuracy (Dice Similarity Coefficient > 0.81 for large structures).
• ResEncM demonstrated significantly lower epistemic uncertainty and ensemble variance across all structures.
• ResEncM showed significantly reduced ECE for the optic chiasm, optic tracts, and pituitary.

Conclusions:

• Integrating a deep residual encoder enhances the reliability and calibration of automated brain OOI contours.
• The ResEncM architecture provides a more trustworthy tool for clinical radiotherapy by reliably flagging high-uncertainty voxels.
• This approach supports confidence-aware clinical workflows in radiotherapy planning.