About model training in LiTS

[Seconight]

Hello🥰 I tried to train the XNet3D on LiTS, the parameters have been checked against the paper, but during the training process, the accuracy of one category is relatively low compared to the other two, such as: (Actually, I have already experimented, but I forgot to modify the samples.per_volume_train in this experiment. But the output of this training is still the same phenomenon) Is that normal?

[Seconight]

and now the results are

which are lower....

[Yanfeng-Zhou]

LiTS is a segmentation data set of liver (1) and liver tumors (2). Compared with the liver, liver tumors are a small target and have variable shapes. Tumor segmentation is relatively difficult. You can visualize some masks to get an intuitive feel.

[Seconight]

Thank you for your reply! I got it. But I encountered a problem in training, and the final result was [0.7270, 0.8101, 54.06, 15.77], which is much lower than that in paper. The best JC during the training process was only 0.6957 lol.

(Additionally, there is no threshold required for inference in the final output...) Is there anything else I haven't noticed about LiTS? How do I need to make corrections?

[Yanfeng-Zhou]

There seems to be no problem with the training process so far. Since the LiTS data set is difficult to converge, I suggest you increase the epoch or reduce the learning rate, and use the training results of the last few epochs for inference.

[Seconight]

It works! Thank you very much for your reply and patience.🤗 Wishing you all the best in your research and success in all your endeavors!