the poor performance when i adjust the args.out_size=512

[ShChen233]

Previously, I used the code from the Medical-SAM-Adapter project (https://github.com/KidsWithTokens/Medical-SAM-Adapter) to perform multi-organ segmentation on the Synapse dataset, achieving an average Dice accuracy of 89%. However, after adjusting related parameters in the finetune-SAM project (such as setting out_size=512) and training using the adapter method, I found a significant decrease in the Dice score. Can you provide some advice? Thank you very much.

[Guhanxue]

aha, May i ask if you adjust your output channel of the network?

Also, It might bc current version of dataset/dataloader have not support multi-class segmentation version.

I was using two different dataset.py before but i thought it was not very organized and elegant. Thus i am planning to combine all my dataset related codes into one Class. i am still organize that part of code (i am so sorry for the inconvience). Do you mind giving me around 1 day to update my repo, i will send you an msg when it is fixed.

[ShChen233]

Thank you very much for your detailed explanation. I have made an adjustment to the output channel of the network by setting the num_classes=args.n_classes. Furthermore, I have recently embarked on a research project involving the adaptation of SAM to multi-organ segmentation. If it's convenient for you, could we possibly exchange contact information, such as WeChat or Email?hhh. Perhaps we can collaborate in the future~

[ShChen233]

[Guhanxue]

cooooool, what is your wechat?

[ShChen233]

emmm，It may not be convenient to show wechat id under the Issues? hhh. If it's convenient for you, could you provide your email address? My email is [email protected]

[Guhanxue]

Hi, i updated the support for multi-cls segmentation. Please free feel to test it.

[ShChen233]

I'm sorry, I forgot that I had updated my email address... It should be [email protected]. @Guhanxue 。And thank you very much for your detailed explanation again~