asa008/a-variation-of-dice-coefficient-loss-caffe-layer

Compute the variation of dice coefficient loss for real-value regression task.

A Variation of Dice Coefficient Loss Layer

If you are interested in Unet or Segmentation, please jump to the Related works.  

Target:

Compute the variation of dice coefficient loss for real-value regression task, such as super resolution. Mathematically,

Loss = (2x^Ty+e)/(x^Tx+y^Ty+e)

where x,y are both vectors in float32. e referes to smooth term (default 1).

How to use?

layer {
  name: "loss"
  type: "DiceCoefLoss"
  bottom: "Deconv"
  bottom: "label"
  top: "loss"
}

The usage is the same as EuclideanLoss layer, restricted to bottom_size==2.

Related works:

Unet Dice Loss for segmentation:

Dice(A,B) = 2|AB|/(|A|+|B|)

Source Code:

1. Caffe Implementation I.
2. Caffe Implementation II.
3. TensorFlow Unet with Dice Loss.