search

JunMa11 / SOTA-MedSeg

SOTA medical image segmentation methods based on various challenges
1.53k stars 200 forks source link

Question of LiTS 2017 dataset processing and performance report? #7

Closed sherjy closed 3 years ago

sherjy commented 3 years ago

Hiiii Jun, thanks for your nice leaderboard collecting repo! Here I have a question for LiTS 2017 dataset performance: https://competitions.codalab.org/competitions/17094#results I noticed in your paper "How Distance Transform Maps Boost Segmentation CNNs: An Empirical Study" you used this dataset for loss comparison. I'm so confused why the performance could vary so far, just like on leaderboard, the liver tumour segmentation result could be up to 0.7x and the results mentioned in your paper is 0.5x. Thougn there may exist some difference on dataset split and processing. image image

I really want to use some template dataset processing codes for this dataset to get 0.7x on baselines like V-Net. Could your give some more guidance? Thanks a lot in advance!

JunMa11 commented 3 years ago

Hi @sherjy ，

The goal of SegDistMap is to compare different distance map usage rather than pursuing SOTA performance. So, we only use a very very simple V-Net.

If you opt to obtain better performance, nnU-Net is definitely worth trying! https://github.com/MIC-DKFZ/nnUNet

Hope the answer helps:)

Best, Jun

sherjy commented 3 years ago

Hi @sherjy ，

The goal of SegDistMap is to compare different distance map usage rather than pursuing SOTA performance. So, we only use a very very simple V-Net.

If you opt to obtain better performance, nnU-Net is definitely worth trying! https://github.com/MIC-DKFZ/nnUNet

Hope the answer helps:)

Best, Jun

Thanks a lot for your advice!