First of all thanks for your great work tony, I really appreciate it.
So far i've trained the net with patch_size (64,64,64) and batch_size of 4 on your provided train.py file due to graphics card limitations. Subsequently i've executed the evaluate_model.py file in order to test the net on your provided test data of 37 volumes. The train, validata as well as the test data all are volumes of dimensions 240 240 155 after a certain MRI channel has been picked.
However, so far i was not able to generate prediction volumes that meet the expected standards. In the evaluate_model.py file i fail to find out how the patch_size works on this model. I've tested the following parameters without success:
"dim": (240,240,155) yields to an error regarding dimensions (I will provide the detailed error message if neccessary)
"dim": (64,64,64) yields to a dice coefficient that meets the expected standards. However, the prediction volume file is cropped to dimensions 64,64,64
"dim": (144,144,144) yields to a dice coefficient that is far below (<0.2) the expected value. The prediction volume file has dimensions 144,144,144
I have not altered the remaining model parameters.
Can you tell me how to create prediction volume files that do have the expected dice coefficient (>0.8 on some volume files) but without cropping the test volume file? I would like the prediction volume file to have dimensions 240,240,155, just like the corresponding test volume file.
First of all thanks for your great work tony, I really appreciate it.
So far i've trained the net with patch_size (64,64,64) and batch_size of 4 on your provided train.py file due to graphics card limitations. Subsequently i've executed the evaluate_model.py file in order to test the net on your provided test data of 37 volumes. The train, validata as well as the test data all are volumes of dimensions 240 240 155 after a certain MRI channel has been picked.
However, so far i was not able to generate prediction volumes that meet the expected standards. In the evaluate_model.py file i fail to find out how the patch_size works on this model. I've tested the following parameters without success:
"dim": (240,240,155) yields to an error regarding dimensions (I will provide the detailed error message if neccessary)
"dim": (64,64,64) yields to a dice coefficient that meets the expected standards. However, the prediction volume file is cropped to dimensions 64,64,64
"dim": (144,144,144) yields to a dice coefficient that is far below (<0.2) the expected value. The prediction volume file has dimensions 144,144,144
I have not altered the remaining model parameters.
Can you tell me how to create prediction volume files that do have the expected dice coefficient (>0.8 on some volume files) but without cropping the test volume file? I would like the prediction volume file to have dimensions 240,240,155, just like the corresponding test volume file.
Best regards