Hi i’m trainig a 2d unet for pancreas segmentation and i found a problem. when i start the training with 128x128 images of the ct scan the weights that are trained have a loss of 0.11 but a dice of 0.55 and i get a prediction similar to my y if I want to see it. after more trainign and scaling image from data to 512x512 (original res) i found a similar loss but when i traied to make the prediction it doesn’t show anything as it predicts that its all Void and no Pancreas.
Anyone has an idea of what should be the problem.
thanks
I have found out that Resnet seems capable of finding the pancreas in a 64x64 and 128x128 CT but it can’t find it over 256x256 resolution.
I have tried unfreeze so it can learn new features at early layers but it doesn’t help.
Anyone has any idea that can help.
Also is there any way to modify Resnet to add some dilation to the kernels. I think that dilation should help compensate a bigger size.
Again thanks for any help you can give me.