Quantifying U-Net uncertainty in multi-parametric MRI-based glioma segmentation by spherical image projection.

Area of Science:

• Medical image analysis
• Deep learning for neuro-oncology
• Uncertainty quantification in AI

Background:

• Quantifying uncertainty in deep learning is crucial for reliable medical image segmentation.
• Current methods lack robust uncertainty measurements, hindering clinical trust.
• Novel approaches are needed to assess the confidence of deep learning segmentation models.

Purpose of the Study:

• To develop a U-Net based method for quantifying segmentation uncertainty in glioma using spherical projection of multi-parametric MRI (MP-MRI).
• To evaluate the proposed method's ability to provide pixel-wise uncertainty maps and scores for improved clinical decision-making.

Main Methods:

• Implemented a Spherical Projection-based U-Net (SPU-Net) model incorporating a nonlinear image transformation for global anatomical information retention.
• Generated multiple independent segmentation predictions from single MP-MRI scans, averaging results to create uncertainty maps.
• Compared SPU-Net against classic U-Net with test-time augmentation (TTA) and linear scaling-based U-Net (LSU-Net) using 369 glioma patient scans.

Main Results:

• SPU-Net demonstrated low uncertainty for correct segmentations and high uncertainty for incorrect predictions, effectively highlighting segmentation errors.
• Achieved superior uncertainty scores (0.826/0.848/0.936 for ET/TC/WT) compared to U-Net with TTA (0.784/0.643/0.872) and LSU-Net (0.743/0.702/0.876).
• SPU-Net also yielded statistically significant improvements in segmentation accuracy, indicated by higher Dice coefficients.

Conclusions:

• The SPU-Net model provides a robust tool for quantifying glioma segmentation uncertainty and enhancing segmentation accuracy.
• This method shows potential for generalization to other medical imaging deep learning applications requiring uncertainty evaluation.