Boundary-weighted logit consistency improves calibration of segmentation networks
This improves calibration for medical image segmentation, addressing a critical issue for reliable predictions in healthcare applications, though it is an incremental advancement over existing regularization methods.
The paper tackled the problem of miscalibration in segmentation networks due to label ambiguity, showing that a boundary-weighted logit consistency regularizer achieves state-of-the-art calibration, with top results on prostate and heart MRI segmentation benchmarks.
Neural network prediction probabilities and accuracy are often only weakly-correlated. Inherent label ambiguity in training data for image segmentation aggravates such miscalibration. We show that logit consistency across stochastic transformations acts as a spatially varying regularizer that prevents overconfident predictions at pixels with ambiguous labels. Our boundary-weighted extension of this regularizer provides state-of-the-art calibration for prostate and heart MRI segmentation.