Training performance issues on small-sized targets

[ZYH23333]

Hello! Thank you for your excellent work!

I encountered a problem while training my dataset using TransBTS. My dataset is about 3D imaging of pulmonary nodules, but the proportion of pulmonary nodules is very small. In a 512x512x150 input volume, the average size of pulmonary nodules is less than 10x10x10.

During two rounds of training with TransBTS, the loss function did not decrease, and none of the predictions segmented the pulmonary nodules. Is it because of the severe imbalance in the ratio of foreground to background, leading to this problem? Do you have any other ideas?

Thank you for your answer!