How to adapt the code to do segmentation of multiple organs?

[xychen2022]

Hello,

I am running the code on multiple partially labelled datasets, including BTCV and FLARE22, with the goal to segment multiple organs using the same model. Since each dataset only have a subset of categories annotated, so I think your method suits this situation well. These datasets are slightly different from the datasets you used in the paper, as they have more than 2 organs annotated for each image. Therefore, I changed the data loader and made sure that in each iteration, only one FG class is selected and the task_id is the same as the corresponding class label. I expected the model to work well but actually I found the model converges slowly and after more than 150 epochs (I used an alternative way to do the same task and after training for this number of epochs, the model can already have reasonable performance), a majority of organs were still not correctly segmented (some even were predicted as all 0s). Could you please share your experiences on how to adapt the code to do segmentation of multiple organs to maximize its performance? And is it expected to have a slow convergence rate when using this method? Thanks!

[Jingnan-Jia]

@xychen2022 Hi bro, liston to me, the repository could not be reproduced at all. Give up it. The authors are not responsible enough. I tried to reproduced their results of the paper several times but I failed. I do not know what happed. This repository wasted me more than 1 month. Just do not have any expectation on the code.

[lphilomena]

Their version 1 (DoDNet /a_DynConv/) works well. I obtained mean dice 0.745 on this version with their code.