用一句英文概括以下所讲的主要内容： The evaluation of medical image segmentation model prediction has traditionally focused on the accuracy index eg Dice coefficient and HD distance which have served as the basis for m

The main content discussed is the evaluation of medical image segmentation model prediction, with a focus on the accuracy index traditionally used for model improvement. However, recent research highlights a reduction in reliability despite improved accuracy. As a result, there is increased attention on model calibration and the reliability of confidence estimates. Different methods, such as dropout-based and Bayesian-based approaches, are commonly used for confidence estimation. Model integration is suggested to improve the reliability of predictive confidence estimation. Additionally, the evaluation of medical AI models should consider both prediction correctness and confidence reliability. Evaluation metrics like expected calibration error (ECE), maximum calibration error (MCE), and Brier score are commonly used. However, the measure of correctness-confidence rank correlation and usable region estimation (URE) are proposed to assess the consistency of rank in confidence estimates.