Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory & Practice
dice 系数和jaccard 索引常常作为医学图像分割任务中指标,在卷积神经网络的训练中常使用加权的交叉熵来作为损失函数,但是在优化的时候两者之间就会存在一种不利差异,交叉熵作用于单像素点,但分割的结果是一个整体。本文主要讨论两个点。探究损失最小化框架中的理论差异,并讨论加权的交叉熵函数只能取代dice 或者 jaccard loss
用上述的loss在5个不同的数据集上测试和验证效果
最终的结果是Through the application of relative approximation bounds, we show that all surrogates are equivalent up to a multiplicative factor, and that no optimal weighting of cross-entropy exists to approximate Dice or Jaccard measures.
指标替代型loss(dice/jaccard)之间是等价的,但是不存在最优近似的加权交叉熵去替代指标替代型lossThey further note that the Dice score reflects both size and localization agreement, more in line with perceptual quality compared to pixel-wise accuracy.
文中提到对比逐像素点的准确率而言,在感知层面上dice 系数在标签不均衡的情况下更能反映尺寸和定位的认同度We are able to validate our finndings empirically on five medical tasks, findin