Issue with Training Phase - Only Achieving 0.2 DSC on Validation Set
Marceloxo opened this issue · comments
I've noticed that during the training phase, the Dice Similarity Coefficient (DSC) consistently remains around 0.2. Is this behavior normal for ModeT? I am working with the OASIS dataset, and my Dataloader is consistent with TransMorph, with img_size = (160, 192, 224). Besides these settings, I haven't made any other modifications. To rule out any environment issues due to PyTorch versions, I tested with torch=1.7.0, 1.9.0, 1.10.0, and 1.11.0, but the results were consistently around 0.2. It's worth mentioning that I previously ran TransMorph successfully using version 1.9.0.
Do you have any insights into why ModeT might exhibit this behavior? I would greatly appreciate your guidance in resolving this issue.
Thank you for your time and assistance.
I wanted to provide additional details regarding my use of the OASIS dataset with TransMorph. I have utilized the provided pkl format files from TransMorph, and for quick testing purposes, I selected 30 volumes for training and 10 for validation. In previous experiments with TransMorph using 300 images, I achieved a DSC of approximately 0.8.
Given these variations in dataset size and performance, I'm seeking insights into why, with the current configuration, ModeT is consistently yielding a DSC of around 0.2 during the training phase.
Your assistance in understanding and resolving this discrepancy would be highly appreciated.
First of all, thank you for your attention to this work. First I want to confirm whether the convergence behavior of ModeT and TransMorph are similar. If there is similar convergence behavior, please confirm again whether the same dice calculation code is used, including the Seg_norm class in data/trans.py and the dsc_val_VOI function in utils.py.
Thank you very much for your advice! You were absolutely right—I indeed forgot to update the segmentation labels corresponding to the dataset. After correcting the calculations, the DSC values now appear to be normal. I must say, ModeT is truly a fantastic registration framework.
I appreciate your prompt assistance and guidance. If there's anything else I should be aware of or if you have further recommendations, please feel free to let me know.
Once again, thank you for your help.
Best regards.