Tumors detected but the segmentation area is way smaller than it should be

[hadign20]

Hello,

I have trained nnUNet on a PNET dataset. Everything works well, but there are two issues when I test the trained model on a validation dataset:

1. Some tumors are not detected at all
2. Some tumors are detected, but the segmentation is way smaller than the ground truth

Please let me know how can I increase the detection rate and make the resulting segmentations larger. If there is any part of the code to change, or there are pre/post processing steps I can take.

Thank you

[fitzjalen]

Hello,

I have trained nnUNet on a PNET dataset. Everything works well, but there are two issues when I test the trained model on a validation dataset:

1. Some tumors are not detected at all

2. Some tumors are detected, but the segmentation is way smaller than the ground truth

Please let me know how can I increase the detection rate and make the resulting segmentations larger.

If there is any part of the code to change, or there are pre/post processing steps I can take.

Thank you

What kind of planner you use? Default or ResEncM/L/XL

Also patch size could play a role there.

[hadign20]

@kndahl I have tried with both default and ResEncM planners.

Where can I change the patch size?

[fitzjalen]

@kndahl I have tried with both default and ResEncM planners.

Where can I change the patch size?

You can do it in you json plans file BTW what patch size you have?

[hadign20]

@kndahl

For the 3d_fullers configuration, the patch size is:

"patch_size": [
                48,
                320,
                320
            ]

and for the 2d configuration, the patch size is:

"patch_size": [
                512,
                512
            ]

I believe these values are selected by running nnUNetv2_plan_and_preprocess. I am using the 3d_fullers configuration. How should I change these values to make the model more sensitive to tumors, so that the segmentations become larger?

plans.json