dice学习

$\left|X\right|$表示$X$的元素个数

Sørensen–Dice coefficient

给定两个集合，则dice coefficient
$$DSC=\frac{2\left|X\cap Y\right|}{\left|X\right|+\left|Y\right|}$$
如果两个集合完全一样，则dice coefficient为1

如果两个集合都是二值（只有0和1），则

$$DSC=\frac{2TP}{2TP+FP+FN}$$
从这也能看出，Dice介于0-1之间

如果是两个向量，则
$$DSC=\frac{2\mathbf{p}\cdot \mathbf{q}}{\mathbf{p}\cdot \mathbf{p}+\mathbf{q}\cdot \mathbf{q}}$$

求导
$$\frac{\partial DSC}{\partial \mathbf{p}}=-2\frac{\mathbf{q}\left(\mathbf{p}\cdot \mathbf{p}+\mathbf{q}\cdot \mathbf{q}\right)-2\mathbf{p}\left(\mathbf{p}\cdot \mathbf{q}\right)}{{\left(\mathbf{p}\cdot \mathbf{p}+\mathbf{q}\cdot \mathbf{q}\right)}^2}$$

DSC与JAC（IOU)的关系
$$JAC=\frac{\left|X\cap Y\right|}{\left|X\cup Y\right|}=\frac{\left|X\cap Y\right|}{\left|X\right|+\left|Y\right|-\left|X\cap Y\right|}$$
因此
$$JAC=\frac{DSC}{2-DSC}$$
所以当dice增大，jac也会增大（如果你看到论文里dice上升了，但是jac没有…

Dice loss

$$DiceLoss=1-DSC$$
代码
这个代码是多个类别的
计算的时候相当于每个类别算一次

def my_dice(pred, target, smooth=1e-5, eps=1e-5):
    """

    :param pred: prediction (BCHW/ BCHWD)
    :param target: target(BHW/ BHWD)
    :param smooth: 
    :param eps: prevent divide by zero
    :return: dice loss
    """
    C = pred.size(1)

    # B1HW/ B1HWD
    target = target.unsqueeze(1)

    sh = list(target.shape)
    sh[1] = C
    o = torch.zeros(size=sh, dtype=target.dtype, device=target.device)
    # one-hot
    target = o.scatter_(dim=1, index=target.long(), value=1)

    reduce_axis = list(range(2, len(pred.shape)))
    intersection = torch.sum(target * pred, dim=reduce_axis)

    ground_o = torch.sum(target, dim=reduce_axis)
    pred_o = torch.sum(pred, dim=reduce_axis)

    denominator = ground_o + pred_o

    result = 1.0 - (2.0 * intersection + smooth) / (denominator + eps)

    result = torch.mean(result)

    return result

也可以用monai

from monai.losses import DiceLoss

Generalized Dice Loss

$$GDL=1-2\frac{{\sum }_{l=1}^2{w}_l{\sum }_n{r}_{ln}{p}_{ln}}{{\sum }_{l=1}^2{w}_l{\sum }_n\left(r_{ln}+p_{ln}\right)}$$
其中$r$是真实值，$p$是预测
$l$是类别，$n$是像素索引，$w_l=\frac{1}{{\left({\sum }_n{r}_{ln}\right)}^2}$

代码

def my_generalized_dice(pred, target, smooth=1e-5, eps=1e-5):
    """

    :param pred: prediction (BCHW/ BCHWD)
    :param target: target(BHW/ BHWD)
    :param smooth:
    :param eps: prevent divide by zero
    :return: dice loss
    """
    C = pred.size(1)

    # B1HW/ B1HWD
    target = target.unsqueeze(1)

    sh = list(target.shape)
    sh[1] = C
    o = torch.zeros(size=sh, dtype=target.dtype, device=target.device)
    # one-hot
    target = o.scatter_(dim=1, index=target.long(), value=1)

    reduce_axis = list(range(2, len(pred.shape)))
    intersection = torch.sum(target * pred, dim=reduce_axis)

    ground_o = torch.sum(target, dim=reduce_axis)
    pred_o = torch.sum(pred, dim=reduce_axis)

    denominator = ground_o + pred_o

    w = 1 / (ground_o * ground_o)
    infs = torch.isinf(w)

    # prevent inf
    w[infs] = 0.0
    max_values = torch.max(w, dim=1)[0].unsqueeze(dim=1)
    w = w + infs * max_values

    numer = 2.0 * torch.sum(intersection * w, dim=1) + smooth
    denom = torch.sum(denominator * w, dim=1) + eps
    result = 1 - numer / denom
    result = torch.mean(result)

    return result

也可以用monai

from monai.losses import GeneralizedDiceLoss

不过感觉monai的有点奇怪，monai是算了每个样本每个类别的，然后一起平均

https://en.wikipedia.org/wiki/S%C3%B8rensen%E2%80%93Dice_coefficient
V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
https://zhuanlan.zhihu.com/p/269592183
Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations