{"id":6475,"date":"2022-08-30T17:18:00","date_gmt":"2022-08-30T09:18:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=6475"},"modified":"2022-09-07T12:06:50","modified_gmt":"2022-09-07T04:06:50","slug":"segment_loss-function","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/08\/30\/segment_loss-function\/","title":{"rendered":"\u56fe\u50cf\u5206\u5272\u635f\u5931\u51fd\u6570loss \u603b\u7ed3+\u4ee3\u7801"},"content":{"rendered":"\n\n\n

\u6c47\u603b\u8bed\u4e49\u5206\u5272\u4e2d\u5e38\u7528\u7684\u635f\u5931\u51fd\u6570\uff1a<\/p>\n\n\n\n

  1. cross entropy loss<\/strong><\/li>
  2. weighted loss<\/strong><\/li>
  3. focal loss<\/strong><\/li>
  4. dice soft loss<\/strong><\/li>
  5. soft iou loss<\/strong><\/li>
  6. Tversky Loss<\/em><\/strong><\/li>
  7. Generalized Dice Loss<\/em><\/strong><\/li>
  8. Boundary Loss<\/em><\/strong><\/li>
  9. Exponential Logarithmic Loss<\/em><\/strong><\/li>
  10. Focal Tversky Loss<\/strong><\/li>
  11. Sensitivity Specificity Loss<\/strong><\/li>
  12. Shape-aware Loss<\/strong><\/li>
  13. Hausdorff Distance Loss<\/strong><\/li><\/ol>\n\n\n\n

    \u53c2\u8003\u8bba\u6587<\/em><\/strong>\uff1aMedical Image Segmentation Using Deep Learning:A Survey<\/a><\/em><\/strong><\/p>\n\n\n\n

    \u8bba\u6587\u5730\u5740<\/strong>:A survey of loss functions for semantic segmentatio<\/a><\/strong><\/em>n<\/em><\/strong>
    \u4ee3\u7801\u5730\u5740<\/strong>\uff1a    https:\/\/github.com\/shruti-jadon\/Semantic-Segmentation-Loss-Functions<\/strong><\/em><\/a>
    \u9879\u76ee\u63a8\u8350<\/strong>\uff1a    https:\/\/github.com\/JunMa11\/SegLoss<\/em><\/strong><\/a><\/p>\n\n\n\n

    \u56fe\u50cf\u5206\u5272\u4e00\u76f4\u662f\u4e00\u4e2a\u6d3b\u8dc3\u7684\u7814\u7a76\u9886\u57df\uff0c\u56e0\u4e3a\u5b83\u6709\u53ef\u80fd\u4fee\u590d\u533b\u7597\u9886\u57df\u7684\u6f0f\u6d1e\uff0c\u5e76\u5e2e\u52a9\u5927\u4f17\u3002\u5728\u8fc7\u53bb\u76845\u5e74\u91cc\uff0c\u5404\u79cd\u8bba\u6587\u63d0\u51fa\u4e86\u4e0d\u540c\u7684\u76ee\u6807\u635f\u5931\u51fd\u6570\uff0c\u7528\u4e8e\u4e0d\u540c\u7684\u60c5\u51b5\u4e0b\uff0c\u5982\u504f\u5dee\u6570\u636e\uff0c\u7a00\u758f\u5206\u5272\u7b49\u3002<\/p>\n\n\n\n

    \u56fe\u50cf\u5206\u5272\u53ef\u4ee5\u5b9a\u4e49\u4e3a\u50cf\u7d20\u7ea7\u522b\u7684\u5206\u7c7b\u4efb\u52a1<\/strong>\u3002\u56fe\u50cf\u7531\u5404\u79cd\u50cf\u7d20\u7ec4\u6210\uff0c\u8fd9\u4e9b\u50cf\u7d20\u7ec4\u5408\u5728\u4e00\u8d77\u5b9a\u4e49\u4e86\u56fe\u50cf\u4e2d\u7684\u4e0d\u540c\u5143\u7d20\uff0c\u56e0\u6b64\u5c06\u8fd9\u4e9b\u50cf\u7d20\u5206\u7c7b\u4e3a\u4e00\u7c7b\u5143\u7d20\u7684\u65b9\u6cd5\u79f0\u4e3a\u8bed\u4e49\u56fe\u50cf\u5206\u5272\u3002\u5728\u8bbe\u8ba1\u57fa\u4e8e\u590d\u6742\u56fe\u50cf\u5206\u5272\u7684\u6df1\u5ea6\u5b66\u4e60\u67b6\u6784\u65f6\uff0c\u901a\u5e38\u4f1a\u9047\u5230\u4e86\u4e00\u4e2a\u81f3\u5173\u91cd\u8981\u7684\u9009\u62e9\uff0c\u5373\u9009\u62e9\u54ea\u4e2a\u635f\u5931\/\u76ee\u6807\u51fd\u6570\uff0c\u56e0\u4e3a\u5b83\u4eec\u4f1a\u6fc0\u53d1\u7b97\u6cd5\u7684\u5b66\u4e60\u8fc7\u7a0b<\/strong>\u3002\u635f\u5931\u51fd\u6570\u7684\u9009\u62e9\u5bf9\u4e8e\u4efb\u4f55\u67b6\u6784\u5b66\u4e60\u6b63\u786e\u7684\u76ee\u6807\u90fd\u662f\u81f3\u5173\u91cd\u8981\u7684\uff0c\u56e0\u6b64\u81ea2012\u5e74\u4ee5\u6765\uff0c\u5404\u79cd\u7814\u7a76\u4eba\u5458\u5f00\u59cb\u8bbe\u8ba1\u9488\u5bf9\u7279\u5b9a\u9886\u57df\u7684\u635f\u5931\u51fd\u6570\uff0c\u4ee5\u4e3a\u5176\u6570\u636e\u96c6\u83b7\u5f97\u66f4\u597d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u8fd9\u4e9b\u635f\u5931\u51fd\u6570\u53ef\u5927\u81f4\u5206\u4e3a4\u7c7b\uff1a\u57fa\u4e8e\u5206\u5e03\u7684\u635f\u5931\u51fd\u6570\uff0c\u57fa\u4e8e\u533a\u57df\u7684\u635f\u5931\u51fd\u6570\uff0c\u57fa\u4e8e\u8fb9\u754c\u7684\u635f\u5931\u51fd\u6570\u548c\u57fa\u4e8e\u590d\u5408\u7684\u635f\u5931\u51fd\u6570\uff08\u00a0Distribution-based,Region-based,\u00a0 Boundary-based,\u00a0 and\u00a0 Compounded\uff09<\/strong>\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    1\u3001cross entropy loss<\/h2>\n\n\n\n

    \u7528\u4e8e\u56fe\u50cf\u8bed\u4e49\u5206\u5272\u4efb\u52a1\u7684\u6700\u5e38\u7528\u635f\u5931\u51fd\u6570\u662f\u50cf\u7d20\u7ea7\u522b\u7684\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u8fd9\u79cd\u635f\u5931\u4f1a\u9010\u4e2a\u68c0\u67e5\u6bcf\u4e2a\u50cf\u7d20\uff0c\u5c06\u5bf9\u6bcf\u4e2a\u50cf\u7d20\u7c7b\u522b\u7684\u9884\u6d4b\u7ed3\u679c\uff08\u6982\u7387\u5206\u5e03\u5411\u91cf\uff09\u4e0e\u6211\u4eec\u7684\u72ec\u70ed\u7f16\u7801\u6807\u7b7e\u5411\u91cf\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n\n\n\n

    \u5047\u8bbe\u6211\u4eec\u9700\u8981\u5bf9\u6bcf\u4e2a\u50cf\u7d20\u7684\u9884\u6d4b\u7c7b\u522b\u67095\u4e2a\uff0c\u5219\u9884\u6d4b\u7684\u6982\u7387\u5206\u5e03\u5411\u91cf\u957f\u5ea6\u4e3a5\uff1a<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    \u6bcf\u4e2a\u50cf\u7d20\u5bf9\u5e94\u7684\u635f\u5931\u51fd\u6570\u4e3a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    \u6574\u4e2a\u56fe\u50cf\u7684\u635f\u5931\u5c31\u662f\u5bf9\u6bcf\u4e2a\u50cf\u7d20\u7684\u635f\u5931\u6c42\u5e73\u5747\u503c\u3002<\/p>\n\n\n\n

    \u7279\u522b\u6ce8\u610f\u7684\u662f\uff0cbinary entropy loss<\/strong> \u662f\u9488\u5bf9\u7c7b\u522b\u53ea\u6709\u4e24\u4e2a\u7684\u60c5\u51b5\uff0c\u7b80\u79f0 bce loss<\/strong>\uff0c\u635f\u5931\u51fd\u6570\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n
    #\u4e8c\u503c\u4ea4\u53c9\u71b5\uff0c\u8fd9\u91cc\u8f93\u5165\u8981\u7ecf\u8fc7sigmoid\u5904\u7406  \nimport torch  \nimport torch.nn as nn  \nimport torch.nn.functional as F  \nnn.BCELoss(F.sigmoid(input), target)  \n#\u591a\u5206\u7c7b\u4ea4\u53c9\u71b5, \u7528\u8fd9\u4e2a loss \u524d\u9762\u4e0d\u9700\u8981\u52a0 Softmax \u5c42  \nnn.CrossEntropyLoss(input, target)<\/code><\/pre>\n\n\n\n
    2\u3001weighted loss<\/h2>\n\n\n\n
    \u7531\u4e8e\u4ea4\u53c9\u71b5\u635f\u5931\u4f1a\u5206\u522b\u8bc4\u4f30\u6bcf\u4e2a\u50cf\u7d20\u7684\u7c7b\u522b\u9884\u6d4b\uff0c\u7136\u540e\u5bf9\u6240\u6709\u50cf\u7d20\u7684\u635f\u5931\u8fdb\u884c\u5e73\u5747\uff0c\u56e0\u6b64\u6211\u4eec\u5b9e\u8d28\u4e0a\u662f\u5728\u5bf9\u56fe\u50cf\u4e2d\u7684\u6bcf\u4e2a\u50cf\u7d20\u8fdb\u884c\u5e73\u7b49\u5730\u5b66\u4e60\u3002\u5982\u679c\u591a\u4e2a\u7c7b\u5728\u56fe\u50cf\u4e2d\u7684\u5206\u5e03\u4e0d\u5747\u8861\uff0c\u90a3\u4e48\u8fd9\u53ef\u80fd\u5bfc\u81f4\u8bad\u7ec3\u8fc7\u7a0b\u7531\u50cf\u7d20\u6570\u91cf\u591a\u7684\u7c7b\u6240\u4e3b\u5bfc\uff0c\u5373\u6a21\u578b\u4f1a\u4e3b\u8981\u5b66\u4e60\u6570\u91cf\u591a\u7684\u7c7b\u522b\u6837\u672c\u7684\u7279\u5f81\uff0c\u5e76\u4e14\u5b66\u4e60\u51fa\u6765\u7684\u6a21\u578b\u4f1a\u66f4\u504f\u5411\u5c06\u50cf\u7d20\u9884\u6d4b\u4e3a\u8be5\u7c7b\u522b\u3002<\/p>\n\n\n\n
    FCN\u8bba\u6587\u548cU-Net\u8bba\u6587\u4e2d\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0c\u5bf9\u8f93\u51fa\u6982\u7387\u5206\u5e03\u5411\u91cf\u4e2d\u7684\u6bcf\u4e2a\u503c\u8fdb\u884c\u52a0\u6743\uff0c\u5373\u5e0c\u671b\u6a21\u578b\u66f4\u52a0\u5173\u6ce8\u6570\u91cf\u8f83\u5c11\u7684\u6837\u672c\uff0c\u4ee5\u7f13\u89e3\u56fe\u50cf\u4e2d\u5b58\u5728\u7684\u7c7b\u522b\u4e0d\u5747\u8861\u95ee\u9898\u3002<\/p>\n\n\n\n
    \u6bd4\u5982\u5bf9\u4e8e\u4e8c\u5206\u7c7b\uff0c\u6b63\u8d1f\u6837\u672c\u6bd4\u4f8b\u4e3a1: 99\uff0c\u6b64\u65f6\u6a21\u578b\u5c06\u6240\u6709\u6837\u672c\u90fd\u9884\u6d4b\u4e3a\u8d1f\u6837\u672c\uff0c\u90a3\u4e48\u51c6\u786e\u7387\u4ecd\u670999%\u8fd9\u4e48\u9ad8\uff0c\u4f46\u5176\u5b9e\u8be5\u6a21\u578b\u6ca1\u6709\u4efb\u4f55\u4f7f\u7528\u4ef7\u503c\u3002<\/p>\n\n\n\n
    \u4e3a\u4e86\u5e73\u8861\u8fd9\u4e2a\u5dee\u8ddd\uff0c\u5c31\u5bf9\u6b63\u6837\u672c\u548c\u8d1f\u6837\u672c\u7684\u635f\u5931\u8d4b\u4e88\u4e0d\u540c\u7684\u6743\u91cd\uff0c\u5e26\u6743\u91cd\u7684\u4e8c\u5206\u7c7b\u635f\u5931\u51fd\u6570\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u8981\u51cf\u5c11\u5047\u9634\u6027\u6837\u672c\u7684\u6570\u91cf\uff0c\u53ef\u4ee5\u589e\u5927 pos_weight\uff1b\u8981\u51cf\u5c11\u5047\u9633\u6027\u6837\u672c\u7684\u6570\u91cf\uff0c\u53ef\u4ee5\u51cf\u5c0f pos_weight\u3002<\/p>\n\n\n\n
    class WeightedCrossEntropyLoss(torch.nn.CrossEntropyLoss):  \r\n   \"\"\"  \r\n   Network has to have NO NONLINEARITY!  \r\n   \"\"\"  \r\n   def __init__(self, weight=None):  \r\n       super(WeightedCrossEntropyLoss, self).__init__()  \r\n       self.weight = weight  \r\n  \r\n   def forward(self, inp, target):  \r\n       target = target.long()  \r\n       num_classes = inp.size()[1]  \r\n  \r\n       i0 = 1  \r\n       i1 = 2  \r\n  \r\n       while i1 < len(inp.shape): # this is ugly but torch only allows to transpose two axes at once  \r\n           inp = inp.transpose(i0, i1)  \r\n           i0 += 1  \r\n           i1 += 1  \r\n  \r\n       inp = inp.contiguous()  \r\n       inp = inp.view(-1, num_classes)  \r\n  \r\n       target = target.view(-1,)  \r\n       wce_loss = torch.nn.CrossEntropyLoss(weight=self.weight)  \r\n  \r\n       return wce_loss(inp, target)<\/code><\/pre>\n\n\n\n
    3\u3001focal loss<\/h2>\n\n\n\n
    \u4e0a\u9762\u9488\u5bf9\u4e0d\u540c\u7c7b\u522b\u7684\u50cf\u7d20\u6570\u91cf\u4e0d\u5747\u8861\u63d0\u51fa\u4e86\u6539\u8fdb\u65b9\u6cd5\uff0c\u4f46\u6709\u65f6\u8fd8\u9700\u8981\u5c06\u50cf\u7d20\u5206\u4e3a\u96be\u5b66\u4e60\u548c\u5bb9\u6613\u5b66\u4e60\u8fd9\u4e24\u79cd\u6837\u672c\u3002<\/p>\n\n\n\n
    \u5bb9\u6613\u5b66\u4e60\u7684\u6837\u672c\u6a21\u578b\u53ef\u4ee5\u5f88\u8f7b\u677e\u5730\u5c06\u5176\u9884\u6d4b\u6b63\u786e\uff0c\u6a21\u578b\u53ea\u8981\u5c06\u5927\u91cf\u5bb9\u6613\u5b66\u4e60\u7684\u6837\u672c\u5206\u7c7b\u6b63\u786e\uff0closs\u5c31\u53ef\u4ee5\u51cf\u5c0f\u5f88\u591a\uff0c\u4ece\u800c\u5bfc\u81f4\u6a21\u578b\u4e0d\u600e\u4e48\u987e\u53ca\u96be\u5b66\u4e60\u7684\u6837\u672c\uff0c\u6240\u4ee5\u6211\u4eec\u8981\u60f3\u529e\u6cd5\u8ba9\u6a21\u578b\u66f4\u52a0\u5173\u6ce8\u96be\u5b66\u4e60\u7684\u6837\u672c\u3002<\/p>\n\n\n\n
    \u5bf9\u4e8e\u8f83\u96be\u5b66\u4e60\u7684\u6837\u672c\uff0c\u5c06 bce loss \u4fee\u6539\u4e3a\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5176\u4e2d\u7684 \u03b3 \u901a\u5e38\u8bbe\u7f6e\u4e3a2\u3002<\/p>\n\n\n\n
    \u901a\u8fc7\u8fd9\u79cd\u4fee\u6539\uff0c\u5c31\u53ef\u4ee5\u4f7f\u6a21\u578b\u66f4\u52a0\u4e13\u6ce8\u4e8e\u5b66\u4e60\u96be\u5b66\u4e60\u7684\u6837\u672c\u3002<\/p>\n\n\n\n
    \u800c\u5c06\u8fd9\u4e2a\u4fee\u6539\u548c\u5bf9\u6b63\u8d1f\u6837\u672c\u4e0d\u5747\u8861\u7684\u4fee\u6539\u5408\u5e76\u5728\u4e00\u8d77\uff0c\u5c31\u662f\u5927\u540d\u9f0e\u9f0e\u7684 focal loss\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    class FocalLoss(nn.Module):  \r\n   \"\"\"  \r\n   copy from: https:\/\/github.com\/Hsuxu\/Loss_ToolBox-PyTorch\/blob\/master\/FocalLoss\/FocalLoss.py  \r\n   This is a implementation of Focal Loss with smooth label cross entropy supported which is proposed in  \r\n   'Focal Loss for Dense Object Detection. (https:\/\/arxiv.org\/abs\/1708.02002)'  \r\n       Focal_Loss= -1*alpha*(1-pt)*log(pt)  \r\n   :param num_class:  \r\n   :param alpha: (tensor) 3D or 4D the scalar factor for this criterion  \r\n   :param gamma: (float,double) gamma > 0 reduces the relative loss for well-classified examples (p>0.5) putting more  \r\n                   focus on hard misclassified example  \r\n   :param smooth: (float,double) smooth value when cross entropy  \r\n   :param balance_index: (int) balance class index, should be specific when alpha is float  \r\n   :param size_average: (bool, optional) By default, the losses are averaged over each loss element in the batch.  \r\n   \"\"\"  \r\n  \r\n   def __init__(self, apply_nonlin=None, alpha=None, gamma=2, balance_index=0, smooth=1e-5, size_average=True):  \r\n       super(FocalLoss, self).__init__()  \r\n       self.apply_nonlin = apply_nonlin  \r\n       self.alpha = alpha  \r\n       self.gamma = gamma  \r\n       self.balance_index = balance_index  \r\n       self.smooth = smooth  \r\n       self.size_average = size_average  \r\n  \r\n       if self.smooth is not None:  \r\n           if self.smooth < 0 or self.smooth > 1.0:  \r\n               raise ValueError('smooth value should be in [0,1]')  \r\n  \r\n   def forward(self, logit, target):  \r\n       if self.apply_nonlin is not None:  \r\n           logit = self.apply_nonlin(logit)  \r\n       num_class = logit.shape[1]  \r\n  \r\n       if logit.dim() > 2:  \r\n           # N,C,d1,d2 -> N,C,m (m=d1*d2*...)  \r\n           logit = logit.view(logit.size(0), logit.size(1), -1)  \r\n           logit = logit.permute(0, 2, 1).contiguous()  \r\n           logit = logit.view(-1, logit.size(-1))  \r\n       target = torch.squeeze(target, 1)  \r\n       target = target.view(-1, 1)  \r\n       # print(logit.shape, target.shape)  \r\n       #   \r\n       alpha = self.alpha  \r\n  \r\n       if alpha is None:  \r\n           alpha = torch.ones(num_class, 1)  \r\n       elif isinstance(alpha, (list, np.ndarray)):  \r\n           assert len(alpha) == num_class  \r\n           alpha = torch.FloatTensor(alpha).view(num_class, 1)  \r\n           alpha = alpha \/ alpha.sum()  \r\n       elif isinstance(alpha, float):  \r\n           alpha = torch.ones(num_class, 1)  \r\n           alpha = alpha * (1 - self.alpha)  \r\n           alpha[self.balance_index] = self.alpha  \r\n  \r\n       else:  \r\n           raise TypeError('Not support alpha type')  \r\n         \r\n       if alpha.device != logit.device:  \r\n           alpha = alpha.to(logit.device)  \r\n  \r\n       idx = target.cpu().long()  \r\n  \r\n       one_hot_key = torch.FloatTensor(target.size(0), num_class).zero_()  \r\n       one_hot_key = one_hot_key.scatter_(1, idx, 1)  \r\n       if one_hot_key.device != logit.device:  \r\n           one_hot_key = one_hot_key.to(logit.device)  \r\n  \r\n       if self.smooth:  \r\n           one_hot_key = torch.clamp(  \r\n               one_hot_key, self.smooth\/(num_class-1), 1.0 - self.smooth)  \r\n       pt = (one_hot_key * logit).sum(1) + self.smooth  \r\n       logpt = pt.log()  \r\n  \r\n       gamma = self.gamma  \r\n  \r\n       alpha = alpha[idx]  \r\n       alpha = torch.squeeze(alpha)  \r\n       loss = -1 * alpha * torch.pow((1 - pt), gamma) * logpt  \r\n  \r\n       if self.size_average:  \r\n           loss = loss.mean()  \r\n       else:  \r\n           loss = loss.sum()  \r\n       return loss<\/code><\/pre>\n\n\n\n
    4\u3001dice soft loss<\/h2>\n\n\n\n
    \u8bed\u4e49\u5206\u5272\u4efb\u52a1\u4e2d\u5e38\u7528\u7684\u8fd8\u6709\u4e00\u4e2a\u57fa\u4e8e Dice \u7cfb\u6570\u7684\u635f\u5931\u51fd\u6570\uff0c\u8be5\u7cfb\u6570\u5b9e\u8d28\u4e0a\u662f\u4e24\u4e2a\u6837\u672c\u4e4b\u95f4\u91cd\u53e0\u7684\u5ea6\u91cf\u3002\u6b64\u5ea6\u91cf\u8303\u56f4\u4e3a 0~1\uff0c\u5176\u4e2d Dice \u7cfb\u6570\u4e3a1\u8868\u793a\u5b8c\u5168\u91cd\u53e0\u3002Dice \u7cfb\u6570\u6700\u521d\u662f\u7528\u4e8e\u4e8c\u8fdb\u5236\u6570\u636e\u7684\uff0c\u53ef\u4ee5\u8ba1\u7b97\u4e3a\uff1a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    |A\u2229B| \u4ee3\u8868\u96c6\u5408A\u548cB\u4e4b\u95f4\u7684\u516c\u5171\u5143\u7d20\uff0c\u5e76\u4e14 |A| \u4ee3\u8868\u96c6\u5408A\u4e2d\u7684\u5143\u7d20\u6570\u91cf\uff08\u5bf9\u4e8e\u96c6\u5408B\u540c\u7406\uff09\u3002<\/p>\n\n\n\n
    \u5bf9\u4e8e\u5728\u9884\u6d4b\u7684\u5206\u5272\u63a9\u7801\u4e0a\u8bc4\u4f30 Dice \u7cfb\u6570\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06 |A\u2229B| \u8fd1\u4f3c\u4e3a\u9884\u6d4b\u63a9\u7801\u548c\u6807\u7b7e\u63a9\u7801\u4e4b\u95f4\u7684\u9010\u5143\u7d20\u4e58\u6cd5\uff0c\u7136\u540e\u5bf9\u7ed3\u679c\u77e9\u9635\u6c42\u548c\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u8ba1\u7b97 Dice \u7cfb\u6570\u7684\u5206\u5b50\u4e2d\u6709\u4e00\u4e2a2\uff0c\u90a3\u662f\u56e0\u4e3a\u5206\u6bcd\u4e2d\u5bf9\u4e24\u4e2a\u96c6\u5408\u7684\u5143\u7d20\u4e2a\u6570\u6c42\u548c\uff0c\u4e24\u4e2a\u96c6\u5408\u7684\u5171\u540c\u5143\u7d20\u88ab\u52a0\u4e86\u4e24\u6b21\u3002 \u4e3a\u4e86\u8bbe\u8ba1\u4e00\u4e2a\u53ef\u4ee5\u6700\u5c0f\u5316\u7684\u635f\u5931\u51fd\u6570\uff0c\u53ef\u4ee5\u7b80\u5355\u5730\u4f7f\u7528 1\u2212Dice\u3002 \u8fd9\u79cd\u635f\u5931\u51fd\u6570\u88ab\u79f0\u4e3a soft Dice loss\uff0c\u8fd9\u662f\u56e0\u4e3a\u6211\u4eec\u76f4\u63a5\u4f7f\u7528\u9884\u6d4b\u51fa\u7684\u6982\u7387\uff0c\u800c\u4e0d\u662f\u4f7f\u7528\u9608\u503c\u5c06\u5176\u8f6c\u6362\u6210\u4e00\u4e2a\u4e8c\u8fdb\u5236\u63a9\u7801\u3002<\/p>\n\n\n\n
    Dice loss\u662f\u9488\u5bf9\u524d\u666f\u6bd4\u4f8b\u592a\u5c0f\u7684\u95ee\u9898\u63d0\u51fa\u7684\uff0cdice\u7cfb\u6570\u6e90\u4e8e\u4e8c\u5206\u7c7b\uff0c\u672c\u8d28\u4e0a\u662f\u8861\u91cf\u4e24\u4e2a\u6837\u672c\u7684\u91cd\u53e0\u90e8\u5206\u3002<\/p>\n\n\n\n
    \u5bf9\u4e8e\u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u51fa\uff0c\u5206\u5b50\u4e0e\u6211\u4eec\u7684\u9884\u6d4b\u548c\u6807\u7b7e\u4e4b\u95f4\u7684\u5171\u540c\u6fc0\u6d3b\u6709\u5173\uff0c\u800c\u5206\u6bcd\u5206\u522b\u4e0e\u6bcf\u4e2a\u63a9\u7801\u4e2d\u7684\u6fc0\u6d3b\u6570\u91cf\u6709\u5173\uff0c\u8fd9\u5177\u6709\u6839\u636e\u6807\u7b7e\u63a9\u7801\u7684\u5c3a\u5bf8\u5bf9\u635f\u5931\u8fdb\u884c\u5f52\u4e00\u5316\u7684\u6548\u679c\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5bf9\u4e8e\u6bcf\u4e2a\u7c7b\u522b\u7684mask\uff0c\u90fd\u8ba1\u7b97\u4e00\u4e2a Dice \u635f\u5931\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5c06\u6bcf\u4e2a\u7c7b\u7684 Dice \u635f\u5931\u6c42\u548c\u53d6\u5e73\u5747\uff0c\u5f97\u5230\u6700\u540e\u7684 Dice soft loss\u3002<\/p>\n\n\n\n
    \u4e0b\u9762\u662f\u4ee3\u7801\u5b9e\u73b0\uff1a<\/p>\n\n\n\n
    def soft_dice_loss(y_true, y_pred, epsilon=1e-6): \n    ''' \n    Soft dice loss calculation for arbitrary batch size, number of classes, and number of spatial dimensions.\n    Assumes the `channels_last` format.\n  \n    # Arguments\n        y_true: b x X x Y( x Z...) x c One hot encoding of ground truth\n        y_pred: b x X x Y( x Z...) x c Network output, must sum to 1 over c channel (such as after softmax) \n        epsilon: Used for numerical stability to avoid divide by zero errors\n    \n    # References\n        V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation \n        https://arxiv.org\/abs\/1606.04797\n        More details on Dice loss formulation \n        https:\/\/mediatum.ub.tum.de\/doc\/1395260\/1395260.pdf (page 72)\n        \n        Adapted from https:\/\/github.com\/Lasagne\/Recipes\/issues\/99#issuecomment-347775022\n    '''\n    \n    # skip the batch and class axis for calculating Dice score<\/em>\n    axes = tuple(range(1, len(y_pred.shape)-1)) \n    numerator = 2. * np.sum(y_pred * y_true, axes)\n    denominator = np.sum(np.square(y_pred) + np.square(y_true), axes)\n    \n    return 1 - np.mean(numerator \/ (denominator + epsilon)) # average over classe<\/em>s and batch<\/code><\/pre>\n\n\n\n
    5\u3001soft IoU loss<\/h2>\n\n\n\n
    \u524d\u9762\u6211\u4eec\u77e5\u9053\u8ba1\u7b97 Dice \u7cfb\u6570\u7684\u516c\u5f0f\uff0c\u5176\u5b9e\u4e5f\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5176\u4e2d TP \u4e3a\u771f\u9633\u6027\u6837\u672c\uff0cFP \u4e3a\u5047\u9633\u6027\u6837\u672c\uff0cFN \u4e3a\u5047\u9634\u6027\u6837\u672c\u3002\u5206\u5b50\u548c\u5206\u6bcd\u4e2d\u7684 TP \u6837\u672c\u90fd\u52a0\u4e86\u4e24\u6b21\u3002<\/p>\n\n\n\n
    IoU \u7684\u8ba1\u7b97\u516c\u5f0f\u548c\u8fd9\u4e2a\u5f88\u50cf\uff0c\u533a\u522b\u5c31\u662f TP \u53ea\u8ba1\u7b97\u4e00\u6b21\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u548c Dice soft loss \u4e00\u6837\uff0c\u901a\u8fc7 IoU \u8ba1\u7b97\u635f\u5931\u4e5f\u662f\u4f7f\u7528\u9884\u6d4b\u7684\u6982\u7387\u503c\uff1a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \u5176\u4e2d C \u8868\u793a\u603b\u7684\u7c7b\u522b\u6570\u3002<\/p>\n\n\n\n
     6\u3001Tversky Loss<\/em><\/strong><\/h2>\n\n\n\n
    \u8bba\u6587\u5730\u5740\u4e3a\uff1ahttps:\/\/arxiv.org\/pdf\/1706.05…<\/a>\u00a0<\/p>\n\n\n\n
          \u533b\u5b66\u5f71\u50cf\u4e2d\u5b58\u5728\u5f88\u591a\u7684\u6570\u636e\u4e0d\u5e73\u8861\u73b0\u8c61\uff0c\u4f7f\u7528\u4e0d\u5e73\u8861\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\u4f1a\u5bfc\u81f4\u4e25\u91cd\u504f\u5411\u9ad8\u7cbe\u5ea6\u4f46\u4f4e\u53ec\u56de\u7387\uff08sensitivity\uff09\u7684\u9884\u6d4b\uff0c\u8fd9\u662f\u4e0d\u5e0c\u671b\u7684\uff0c\u7279\u522b\u662f\u5728\u533b\u5b66\u5e94\u7528\u4e2d\uff0c\u5047\u9634\u6027\u6bd4\u5047\u9633\u6027\u66f4\u96be\u5bb9\u5fcd\u3002\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8eTversky\u6307\u6570\u7684\u5e7f\u4e49\u635f\u5931\u51fd\u6570\uff0c\u89e3\u51b3\u4e86\u4e09\u7ef4\u5168\u5377\u79ef\u6df1\u795e\u7ecf\u7f51\u7edc\u8bad\u7ec3\u4e2d\u6570\u636e\u4e0d\u5e73\u8861\u7684\u95ee\u9898\uff0c\u5728\u7cbe\u5ea6\u548c\u53ec\u56de\u7387\u4e4b\u95f4\u53d6\u5f97\u4e86\u8f83\u597d\u7684\u6298\u8877\u3002<\/p>\n\n\n\n
    Dice loss\u7684\u6b63\u5219\u5316\u7248\u672c\uff0c\u4ee5\u63a7\u5236\u5047\u9633\u6027\u548c\u5047\u9634\u6027\u5bf9\u635f\u5931\u51fd\u6570\u7684\u8d21\u732e\uff0cTL\u88ab\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    class TverskyLoss(nn.Module):  \r\n   def __init__(self, apply_nonlin=None, batch_dice=False, do_bg=True, smooth=1.,  \r\n                square=False):  \r\n       \"\"\"  \r\n       paper: https:\/\/arxiv.org\/pdf\/1706.05721.pdf  \r\n       \"\"\"  \r\n       super(TverskyLoss, self).__init__()  \r\n  \r\n       self.square = square  \r\n       self.do_bg = do_bg  \r\n       self.batch_dice = batch_dice  \r\n       self.apply_nonlin = apply_nonlin  \r\n       self.smooth = smooth  \r\n       self.alpha = 0.3  \r\n       self.beta = 0.7  \r\n  \r\n   def forward(self, x, y, loss_mask=None):  \r\n       shp_x = x.shape  \r\n  \r\n       if self.batch_dice:  \r\n           axes = [0] + list(range(2, len(shp_x)))  \r\n       else:  \r\n           axes = list(range(2, len(shp_x)))  \r\n  \r\n       if self.apply_nonlin is not None:  \r\n           x = self.apply_nonlin(x)  \r\n  \r\n       tp, fp, fn = get_tp_fp_fn(x, y, axes, loss_mask, self.square)  \r\n  \r\n  \r\n       tversky = (tp + self.smooth) \/ (tp + self.alpha*fp + self.beta*fn + self.smooth)  \r\n  \r\n       if not self.do_bg:  \r\n           if self.batch_dice:  \r\n               tversky = tversky[1:]  \r\n           else:  \r\n               tversky = tversky[:, 1:]  \r\n       tversky = tversky.mean()  \r\n  \r\n       return -tversky  <\/code><\/pre>\n\n\n\n
    7\u3001Generalized Dice Loss<\/em><\/strong><\/h2>\n\n\n\n
    Dice loss\u867d\u7136\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u89e3\u51b3\u4e86\u5206\u7c7b\u5931\u8861\u7684\u95ee\u9898\uff0c\u4f46\u5374\u4e0d\u5229\u4e8e\u4e25\u91cd\u7684\u5206\u7c7b\u4e0d\u5e73\u8861\u3002\u4f8b\u5982\u5c0f\u76ee\u6807\u5b58\u5728\u4e00\u4e9b\u50cf\u7d20\u7684\u9884\u6d4b\u8bef\u5dee\uff0c\u8fd9\u5f88\u5bb9\u6613\u5bfc\u81f4Dice\u7684\u503c\u53d1\u751f\u5f88\u5927\u7684\u53d8\u5316\u3002Sudre\u7b49\u4eba\u63d0\u51fa\u4e86Generalized Dice Loss (GDL)<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
        GDL\u4f18\u4e8eDice\u635f\u5931\uff0c\u56e0\u4e3a\u4e0d\u540c\u7684\u533a\u57df\u5bf9\u635f\u5931\u6709\u76f8\u4f3c\u7684\u8d21\u732e\uff0c\u5e76\u4e14GDL\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u66f4\u7a33\u5b9a\u548c\u9c81\u68d2\u3002<\/p>\n\n\n\n
    8\u3001Boundary Loss<\/em><\/strong><\/h2>\n\n\n\n
        \u4e3a\u4e86\u89e3\u51b3\u7c7b\u522b\u4e0d\u5e73\u8861\u7684\u95ee\u9898\uff0cKervadec\u7b49\u4eba[95]\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u7528\u4e8e\u8111\u635f\u4f24\u5206\u5272\u7684\u8fb9\u754c\u635f\u5931\u3002\u8be5\u635f\u5931\u51fd\u6570\u65e8\u5728\u6700\u5c0f\u5316\u5206\u5272\u8fb9\u754c\u548c\u6807\u8bb0\u8fb9\u754c\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002\u4f5c\u8005\u5728\u4e24\u4e2a\u6ca1\u6709\u6807\u7b7e\u7684\u4e0d\u5e73\u8861\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u7ed3\u679c\u8868\u660e\uff0cDice los\u548cBoundary los\u7684\u7ec4\u5408\u4f18\u4e8e\u5355\u4e00\u7ec4\u5408\u3002\u590d\u5408\u635f\u5931\u7684\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \u5176\u4e2d\u7b2c\u4e00\u90e8\u5206\u662f\u4e00\u4e2a\u6807\u51c6\u7684Dice los\uff0c\u5b83\u88ab\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \u7b2c\u4e8c\u90e8\u5206\u662fBoundary los\uff0c\u5b83\u88ab\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    9\u3001Exponential Logarithmic Loss<\/em><\/strong><\/h2>\n\n\n\n

    \u5728\uff089\uff09\u4e2d\uff0c\u52a0\u6743Dice los\u5b9e\u9645\u4e0a\u662f\u5f97\u5230\u7684Dice\u503c\u9664\u4ee5\u6bcf\u4e2a\u6807\u7b7e\u7684\u548c\uff0c\u5bf9\u4e0d\u540c\u5c3a\u5ea6\u7684\u5bf9\u8c61\u8fbe\u5230\u5e73\u8861\u3002\u56e0\u6b64\uff0cWong\u7b49\u4eba\u7ed3\u5408focal loss [96] \u548cdice loss\uff0c\u63d0\u51fa\u4e86\u7528\u4e8e\u8111\u5206\u5272\u7684\u6307\u6570\u5bf9\u6570\u635f\u5931(EXP\u635f\u5931)\uff0c\u4ee5\u89e3\u51b3\u4e25\u91cd\u7684\u7c7b\u4e0d\u5e73\u8861\u95ee\u9898\u3002\u901a\u8fc7\u5f15\u5165\u6307\u6570\u5f62\u5f0f\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63a7\u5236\u635f\u5931\u51fd\u6570\u7684\u975e\u7ebf\u6027\uff0c\u4ee5\u63d0\u9ad8\u5206\u5272\u7cbe\u5ea6\u3002EXP\u635f\u5931\u51fd\u6570\u7684\u5b9a\u4e49\u4e3a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5176\u4e2d\uff0c\u4e24\u4e2a\u65b0\u7684\u53c2\u6570\u6743\u91cd\u5206\u522b\u7528\u03c9dice\u548c\u03c9cross\u8868\u793a\u3002Ldice\u662f\u6307\u6570\u5bf9\u6570\u9ab0\u5b50\u635f\u5931\uff0c\u800c\u4ea4\u53c9\u635f\u5931\u662f\u4ea4\u53c9\u71b5\u635f\u5931<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
         \u5176\u4e2dx\u662f\u50cf\u7d20\u4f4d\u7f6e\uff0ci\u662f\u6807\u7b7e\uff0cl\u662f\u4f4d\u7f6ex\u5904\u7684\u5730\u9762\u771f\u503c\u3002pi(x)\u662f\u4ecesoftmax\u8f93\u51fa\u7684\u6982\u7387\u503c\u3002
    \u5728\uff0817\uff09\u4e2d\uff0cfk\u662f\u6807\u7b7ek\u51fa\u73b0\u7684\u9891\u7387\uff0c\u8be5\u53c2\u6570\u53ef\u4ee5\u51cf\u5c11\u66f4\u9891\u7e41\u51fa\u73b0\u7684\u6807\u7b7e\u7684\u5f71\u54cd\u3002\u03b3Dice\u548c\u03b3cross\u90fd\u7528\u4e8e\u589e\u5f3a\u635f\u5931\u51fd\u6570\u7684\u975e\u7ebf\u6027\u3002<\/p>\n\n\n\n
    10.Focal Tversky Loss<\/strong><\/h2>\n\n\n\n
        \u4e0e\u201cFocal loss\u201d\u76f8\u4f3c\uff0c\u540e\u8005\u7740\u91cd\u4e8e\u901a\u8fc7\u964d\u4f4e\u6613\u7528\/\u5e38\u89c1\u635f\u5931\u7684\u6743\u91cd\u6765\u8bf4\u660e\u56f0\u96be\u7684\u4f8b\u5b50\u3002Focal Tversky Loss\u8fd8\u5c1d\u8bd5\u501f\u52a9\u03b3\u7cfb\u6570\u6765\u5b66\u4e60\u8bf8\u5982\u5728ROI\uff08\u611f\u5174\u8da3\u533a\u57df\uff09\u8f83\u5c0f\u7684\u60c5\u51b5\u4e0b\u7684\u56f0\u96be\u793a\u4f8b\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    class FocalTversky_loss(nn.Module):  \n   \"\"\"  \n   paper: https:\/\/arxiv.org\/pdf\/1810.07842.pdf  \n   author code: https:\/\/github.com\/nabsabraham\/focal-tversky-unet\/blob\/347d39117c24540400dfe80d106d2fb06d2b99e1\/losses.py#L65  \n   \"\"\"  \n   def __init__(self, tversky_kwargs, gamma=0.75):  \n       super(FocalTversky_loss, self).__init__()  \n       self.gamma = gamma  \n       self.tversky = TverskyLoss(**tversky_kwargs)  \n  \n   def forward(self, net_output, target):  \n       tversky_loss = 1 + self.tversky(net_output, target) # = 1-tversky(net_output, target)  \n       focal_tversky = torch.pow(tversky_loss, self.gamma)  \n       return focal_tversky  <\/code><\/pre>\n\n\n\n
    11\u3001Sensitivity Specificity Loss<\/strong><\/h2>\n\n\n\n
    \u9996\u5148\u654f\u611f\u6027\u5c31\u662f\u53ec\u56de\u7387\uff0c\u68c0\u6d4b\u51fa\u786e\u5b9e\u6709\u75c5\u7684\u80fd\u529b\uff1a<\/p>\n\n\n\n
    \"640-8.png\"<\/figure><\/div>\n\n\n\n
    \u7279\u5f02\u6027\uff0c\u68c0\u6d4b\u51fa\u786e\u5b9e\u6ca1\u75c5\u7684\u80fd\u529b\uff1a<\/p>\n\n\n\n
    \"640-9.png\"<\/figure><\/div>\n\n\n\n
    \u800cSensitivity Specificity Loss\u4e3a\uff1a<\/p>\n\n\n\n
    \"640-10.png\"<\/figure>\n\n\n\n
    \"image.png\"<\/figure>\n\n\n\n
    class SSLoss(nn.Module):  \r\n   def __init__(self, apply_nonlin=None, batch_dice=False, do_bg=True, smooth=1.,  \r\n                square=False):  \r\n       \"\"\"  \r\n       Sensitivity-Specifity loss  \r\n       paper: http:\/\/www.rogertam.ca\/Brosch_MICCAI_2015.pdf  \r\n       tf code: https:\/\/github.com\/NifTK\/NiftyNet\/blob\/df0f86733357fdc92bbc191c8fec0dcf49aa5499\/niftynet\/layer\/loss_segmentation.py#L392  \r\n       \"\"\"  \r\n       super(SSLoss, self).__init__()  \r\n  \r\n       self.square = square  \r\n       self.do_bg = do_bg  \r\n       self.batch_dice = batch_dice  \r\n       self.apply_nonlin = apply_nonlin  \r\n       self.smooth = smooth  \r\n       self.r = 0.1 # weight parameter in SS paper  \r\n  \r\n   def forward(self, net_output, gt, loss_mask=None):  \r\n       shp_x = net_output.shape  \r\n       shp_y = gt.shape  \r\n       # class_num = shp_x[1]  \r\n         \r\n       with torch.no_grad():  \r\n           if len(shp_x) != len(shp_y):  \r\n               gt = gt.view((shp_y[0], 1, *shp_y[1:]))  \r\n  \r\n           if all([i == j for i, j in zip(net_output.shape, gt.shape)]):  \r\n               # if this is the case then gt is probably already a one hot encoding  \r\n               y_onehot = gt  \r\n           else:  \r\n               gt = gt.long()  \r\n               y_onehot = torch.zeros(shp_x)  \r\n               if net_output.device.type == \"cuda\":  \r\n                   y_onehot = y_onehot.cuda(net_output.device.index)  \r\n               y_onehot.scatter_(1, gt, 1)  \r\n  \r\n       if self.batch_dice:  \r\n           axes = [0] + list(range(2, len(shp_x)))  \r\n       else:  \r\n           axes = list(range(2, len(shp_x)))  \r\n  \r\n       if self.apply_nonlin is not None:  \r\n           softmax_output = self.apply_nonlin(net_output)  \r\n         \r\n       # no object value  \r\n       bg_onehot = 1 - y_onehot  \r\n       squared_error = (y_onehot - softmax_output)**2  \r\n       specificity_part = sum_tensor(squared_error*y_onehot, axes)\/(sum_tensor(y_onehot, axes)+self.smooth)  \r\n       sensitivity_part = sum_tensor(squared_error*bg_onehot, axes)\/(sum_tensor(bg_onehot, axes)+self.smooth)  \r\n  \r\n       ss = self.r * specificity_part + (1-self.r) * sensitivity_part  \r\n  \r\n       if not self.do_bg:  \r\n           if self.batch_dice:  \r\n               ss = ss[1:]  \r\n           else:  \r\n               ss = ss[:, 1:]  \r\n       ss = ss.mean()  \r\n  \r\n       return ss<\/code><\/pre>\n\n\n\n
    12\u3001Log-Cosh Dice Loss<\/strong><\/h2>\n\n\n\n
    Dice\u7cfb\u6570\u662f\u4e00\u79cd\u7528\u4e8e\u8bc4\u4f30\u5206\u5272\u8f93\u51fa\u7684\u5ea6\u91cf\u6807\u51c6\u3002\u5b83\u4e5f\u5df2\u4fee\u6539\u4e3a\u635f\u5931\u51fd\u6570\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u5b9e\u73b0\u5206\u5272\u76ee\u6807\u7684\u6570\u5b66\u8868\u793a\u3002\u4f46\u662f\u7531\u4e8e\u5176\u975e\u51f8\u6027\uff0c\u5b83\u591a\u6b21\u90fd\u65e0\u6cd5\u83b7\u5f97\u6700\u4f73\u7ed3\u679c\u3002Lovsz-softmax\u635f\u5931\u65e8\u5728\u901a\u8fc7\u6dfb\u52a0\u4f7f\u7528Lovsz\u6269\u5c55\u7684\u5e73\u6ed1\u6765\u89e3\u51b3\u975e\u51f8\u635f\u5931\u51fd\u6570\u7684\u95ee\u9898\u3002\u540c\u65f6\uff0cLog-Cosh\u65b9\u6cd5\u5df2\u5e7f\u6cdb\u7528\u4e8e\u57fa\u4e8e\u56de\u5f52\u7684\u95ee\u9898\u4e2d\uff0c\u4ee5\u5e73\u6ed1\u66f2\u7ebf\u3002<\/p>\n\n\n\n
    \"640.png\"<\/figure><\/div>\n\n\n\n
    \"640-1.png\"<\/figure><\/div>\n\n\n\n
    \u5c06Cosh(x)\u51fd\u6570\u548cLog(x)\u51fd\u6570\u5408\u5e76\uff0c\u53ef\u4ee5\u5f97\u5230Log-Cosh Dice Loss\uff1a<\/p>\n\n\n\n
    \"640-2.png\"<\/figure><\/div>\n\n\n\n
    def log_cosh_dice_loss(self, y_true, y_pred):  \n       x = self.dice_loss(y_true, y_pred)  \n       return tf.math.log((torch.exp(x) + torch.exp(-x)) \/ 2.0)  <\/code><\/pre>\n\n\n\n
    13\u3001Hausdorff Distance Loss<\/strong><\/h2>\n\n\n\n
    Hausdorff Distance Loss\uff08HD\uff09\u662f\u5206\u5272\u65b9\u6cd5\u7528\u6765\u8ddf\u8e2a\u6a21\u578b\u6027\u80fd\u7684\u5ea6\u91cf\u3002\u5b83\u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n
    \"640-4.png\"<\/figure><\/div>\n\n\n\n
        \u4efb\u4f55\u5206\u5272\u6a21\u578b\u7684\u76ee\u7684\u90fd\u662f\u4e3a\u4e86\u6700\u5927\u5316Hausdorff\u8ddd\u79bb\uff0c\u4f46\u662f\u7531\u4e8e\u5176\u975e\u51f8\u6027\uff0c\u56e0\u6b64\u5e76\u672a\u5e7f\u6cdb\u7528\u4f5c\u635f\u5931\u51fd\u6570\u3002\u6709\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u57fa\u4e8eHausdorff\u8ddd\u79bb\u7684\u635f\u5931\u51fd\u6570\u76843\u4e2a\u53d8\u91cf\uff0c\u5b83\u4eec\u90fd\u7ed3\u5408\u4e86\u5ea6\u91cf\u7528\u4f8b\uff0c\u5e76\u786e\u4fdd\u635f\u5931\u51fd\u6570\u6613\u4e8e\u5904\u7406\u3002<\/p>\n\n\n\n
    class HDDTBinaryLoss(nn.Module):  \n   def __init__(self):  \n       \"\"\"  \n       compute haudorff loss for binary segmentation  \n       https:\/\/arxiv.org\/pdf\/1904.10030v1.pdf          \n       \"\"\"  \n       super(HDDTBinaryLoss, self).__init__()  \n  \n  \n   def forward(self, net_output, target):  \n       \"\"\"  \n       net_output: (batch_size, 2, x,y,z)  \n       target: ground truth, shape: (batch_size, 1, x,y,z)  \n       \"\"\"  \n       net_output = softmax_helper(net_output)  \n       pc = net_output[:, 1, ...].type(torch.float32)  \n       gt = target[:,0, ...].type(torch.float32)  \n       with torch.no_grad():  \n           pc_dist = compute_edts_forhdloss(pc.cpu().numpy()>0.5)  \n           gt_dist = compute_edts_forhdloss(gt.cpu().numpy()>0.5)  \n       # print('pc_dist.shape: ', pc_dist.shape)  \n         \n       pred_error = (gt - pc)**2  \n       dist = pc_dist**2 + gt_dist**2 # \\alpha=2 in eq(8)  \n  \n       dist = torch.from_numpy(dist)  \n       if dist.device != pred_error.device:  \n           dist = dist.to(pred_error.device).type(torch.float32)  \n  \n       multipled = torch.einsum(\"bxyz,bxyz->bxyz\", pred_error, dist)  \n       hd_loss = multipled.mean()  \n  \n       return hd_loss<\/code><\/pre>\n\n\n\n
    \u603b\u7ed3\uff1a<\/h2>\n\n\n\n
    \u4ea4\u53c9\u71b5\u635f\u5931\u628a\u6bcf\u4e2a\u50cf\u7d20\u90fd\u5f53\u4f5c\u4e00\u4e2a\u72ec\u7acb\u6837\u672c\u8fdb\u884c\u9884\u6d4b\uff0c\u800c dice loss \u548c iou loss \u5219\u4ee5\u4e00\u79cd\u66f4\u201c\u6574\u4f53\u201d\u7684\u65b9\u5f0f\u6765\u770b\u5f85\u6700\u7ec8\u7684\u9884\u6d4b\u8f93\u51fa\u3002<\/p>\n\n\n\n
    \u8fd9\u4e24\u7c7b\u635f\u5931\u662f\u9488\u5bf9\u4e0d\u540c\u60c5\u51b5\uff0c\u5404\u6709\u4f18\u70b9\u548c\u7f3a\u70b9\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u53ef\u4ee5\u540c\u65f6\u4f7f\u7528\u8fd9\u4e24\u7c7b\u635f\u5931\u6765\u8fdb\u884c\u4e92\u8865\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
    \u6c47\u603b\u8bed\u4e49\u5206\u5272\u4e2d\u5e38\u7528\u7684\u635f\u5931\u51fd\u6570\uff1a cross entropy loss weighted loss focal  … \u7ee7\u7eed\u9605\u8bfb\u56fe\u50cf\u5206\u5272\u635f\u5931\u51fd\u6570loss \u603b\u7ed3+\u4ee3\u7801<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6475"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=6475"}],"version-history":[{"count":61,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6475\/revisions"}],"predecessor-version":[{"id":7455,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6475\/revisions\/7455"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=6475"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=6475"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=6475"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}