search

DIAL-RPI / PIPO-FAN

PIPO-FAN for multi organ segmentation over partial labeled datasets using pytorch
MIT License
59 stars 22 forks source link

Target Adaptive Loss #5

Closed kshen3778 closed 3 years ago

kshen3778 commented 3 years ago

Where can we find the implementation of the target adaptive loss mentioned in the paper? Also what is the difference between the concave_dps and concave_dps_w? It seems the Softmax, attention, and pooling only exists in concave_dps_w.py?

xifang001 commented 3 years ago

Hi, Shen,

The target adaptive loss can be seen at the 302nd in train_sf_partial.py. More general version may be delivered in future. Concave_dps corresponds to the PIPO part, the outputs of PIPO are five feature maps, which helps we learn the weights for different scales. On the other hand, concave_dps_w correspond to PIPO + FAN, the output of is a probability map. Thus softmax and attention are added in concave_dps_w. That is to say, concave_dps (PIPO) is part of concave_dps_w (PIPO+FAN).

Best, Xi

kshen3778 commented 3 years ago

Thank you. Sorry, forgot to reply earlier. Also, in the paper it says " For each epoch, we randomly select three continuous slices containing target organ label from all the CT training volumes" is that the number of input channels to the model - the input size would be (batch size, 3, width, height)? Would the target also be the masks for those three slices or only the middle slice?

xifang001 commented 3 years ago

Yes. The input size is set to be (batch size, 3, width, height) to enhance the context information. It's also named as 2.5D network. Only the mask of middle slice is used for training.

kshen3778 commented 3 years ago

That makes sense. Thanks!