Uncertainty quantification of U-Net based segmentation tool using conformal prediction.

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiotherapy

Background:

• Radiation therapy planning is labor-intensive.
• Artificial intelligence (AI) auto-contouring tools improve efficiency.
• Current AI acceptance testing lacks uncertainty assessment.

Purpose of the Study:

• Demonstrate conformal prediction for AI auto-contouring.
• Provide spatially localized uncertainty information.
• Complement traditional error analysis metrics.

Main Methods:

• Trained a U-Net with ResNet-34 encoder on breast cancer CT scans.
• Evaluated AI auto-contouring using IoU and HD95 metrics.
• Applied conformal prediction with adaptive prediction sets to localize uncertainty.

Main Results:

• Achieved high U-Net performance (mean IoU 0.924, mean HD95 11.35).
• Conformal prediction at 90% confidence showed minimal IoU differences (0.89-1.46%).
• IoU derivatives significantly differed between true and false positives (P < 0.001).

Conclusions:

• Conformal prediction is a valuable tool for AI auto-contouring acceptance testing.
• It spatially localizes prediction uncertainty.
• Enables identification of false positives beyond traditional metrics.