CONSeg: Voxelwise Uncertainty Quantification for Glioma Segmentation Using Conformal Prediction

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Accurate glioma segmentation is crucial for clinical decision-making and treatment planning.
• Uncertainty quantification (UQ) methods, such as conformal prediction (CP), enhance the reliability of segmentation models.
• CP provides statistical confidence guarantees for UQ in medical image analysis.

Purpose of the Study:

• To implement conformal prediction (CP) for uncertainty quantification in glioma segmentation.
• To evaluate the reliability and performance of CP-enhanced glioma segmentation models.
• To assess the correlation between uncertainty measures and segmentation accuracy.

Main Methods:

• Utilized UCSF and UPenn glioma datasets for training, validation, calibration, and testing.
• Trained a UNet model and applied prediction normalization with an optimal threshold of 0.5.
• Implemented CP by selecting a conformal threshold based on calibration nonconformity scores.
• Defined an uncertainty ratio (UR) and evaluated its correlation with Dice Score Coefficient (DSC) and Hausdorff Distance 95 (HD95).

Main Results:

• The base model achieved DSC of 0.86 (internal) and 0.83 (external), with HD95 of 7.35 and 11.71, respectively.
• CP demonstrated high coverage (0.9982 internal, 0.9977 external).
• Significant correlations were found between UR and segmentation metrics (DSC, HD95) (p < 0.001).
• Cases categorized as 'certain' showed significantly better segmentation performance than 'uncertain' cases (p < 0.001).

Conclusions:

• Conformal prediction (CP) effectively quantifies uncertainty in glioma segmentation, enhancing model reliability.
• The proposed conformal segmentation (CONSeg) method improves human-computer interaction by identifying uncertain segmentations.
• CONSeg can flag uncertain cases, recommending them for manual segmentation and improving overall clinical workflow.