Soft Dice Confidence: A Near-Optimal Confidence Estimator for Selective Prediction in Semantic Segmentation

Selective prediction augments a model with the option to abstain from providing unreliable predictions. The key ingredient is a confidence score function, which should be directly related to the conditional risk. In the case of binary semantic segmentation, existing score functions either ignore the particularities of the evaluation metric or demand additional held-out data for tuning. We propose the Soft Dice Confidence (SDC), a simple, tuning-free confidence score function that directly aligns with the Dice coefficient metric. We prove that, under conditional independence, the SDC is near optimal: we establish upper and lower bounds on the ratio between the SDC and the ideal (intractable) confidence score function and show that these bounds are very close to 1. Experiments on six public medical-imaging benchmarks and on synthetic data corroborate our theoretical findings. In fact, SDC outperformed all prior confidence estimators from the literature in all of our experiments, including those that rely on additional data. These results position SDC as a reliable and efficient confidence estimator for selective prediction in semantic segmentation.