MedicalNet icon indicating copy to clipboard operation
MedicalNet copied to clipboard

Published 20 hours ago • verified publisher icon Tencent

做肝脏分割一阶段粗分割得到的dice只有65%算正常吗?好像太低了吧

Open Alagirl opened this issue 5 years ago • 7 comments

做肝脏分割一阶段粗分割得到的dice只有65%算正常吗?好像太低了吧,用Unet3d也可以很快的训练达到同样的精度,想问一下你们的实验也是这样吗,还是我中间有错误

Alagirl avatar Sep 25 '19 09:09 Alagirl

求解答

Alagirl avatar Sep 25 '19 09:09 Alagirl

请问你decoder部分用的是什么？demo里提供的decoder比较简单，不同任务建议用pretrain+合适的decoder，我当时做肝脏参考了ASPP。

cshwhale avatar Sep 25 '19 09:09 cshwhale

我是直接转置卷积上采样,因为我看论文里第一阶段就是直接上采样吧,得到ROI再做ASPP的

Alagirl avatar Sep 26 '19 02:09 Alagirl

我这里单做肺部裁剪 得到的效果也不是特别理想 不清楚是不是中间产生什么错误还是必须添加aspp

luofeel avatar Oct 09 '19 07:10 luofeel

请问你decoder部分用的是什么？demo里提供的decoder比较简单，不同任务建议用pretrain+合适的decoder，我当时做肝脏参考了ASPP。

请问是在ASPP后接上采样吗？

joechenrh avatar May 11 '20 08:05 joechenrh

I have the same problem. I trained the LIDC-IDRI segmentation dataset with the MedicalNet pretrained model(resnet 18), and got a dice score lower than traning from scratch(54.8 vs 74.0). Any suggestions?

urmagicsmine avatar Dec 11 '20 02:12 urmagicsmine

Hi @urmagicsmine @joechenrh @luofeel @Alagirl @cshwhale

For liver/liver tumor segmentation, probably you can try this post: https://github.com/MrGiovanni/ModelsGenesis/tree/master/competition

They also used a pre-trained 3D model and seemed to outperform training nnU-Net from scratch in liver/liver tumor.

MDAooo avatar Dec 11 '20 04:12 MDAooo