search

josedolz / SemiDenseNet

Repository containing the code of one of the networks that we employed in the iSEG Grand MICCAI Challenge 2017, infant brain segmentation.
MIT License
55 stars 18 forks source link

About ideas on receptive field #1

Closed CaelynCheung1996 closed 6 years ago

CaelynCheung1996 commented 6 years ago

Hi Jose,

Recently I am running your method on MR segmentation task. I selected a small rectangle region as ROI to segment the target into it. The DICE is not bad but it seems like there's always some fragments (some tiny spots I consider) wrongly segmented, which is obviously not belonged to my target. Is that cause by the small receptive field problem or do you have any idea what is this about? Thanks.

josedolz commented 6 years ago

Hi @CaelynCheung1996

This isolated outliers might be due to the limited receptive field, as you said. If they are not connected to your main target you could remove them with some easy post-processing. A solution for that would be to replace the convolutions by dilated convolutions, so that you increase the receptive field of the net without degrading the performance.

Jose.

CaelynCheung1996 commented 6 years ago

Thanks for your opinion Jose. I think I will try CRF first to see it it works. BTW, when will the latest HyperDenseNet be released?

josedolz commented 6 years ago

I expect to release the code of HyperDenseNet in the following weeks (max 4-5 weeks).

Best,