Gradient Vector Field Surgery for Calibrated Segmentation
July 17, 2026
Gradient vector field surgery mitigates overconfidence in region-based segmentation models by scaling the loss's partial derivative linearly with prediction error. The method improves calibration in 2D and 3D medical segmentation tasks without sacrificing performance.
HOW THIS AFFECTS YOU
●
researcherThis offers a simple gradient-based intervention to fix miscalibration caused by Dice-style losses.
●
healthThis improves the clinical reliability of segmentation models used for defining tumor margins.