2583 Add DatasetSummary #2616

yiheng-wang-nv · 2021-07-16T09:35:42Z

Signed-off-by: Yiheng Wang [email protected]

Fixes #2583 .

Description

For 3D image segmentation tasks, users may not sure how to set a reasonable voxel spacing value, as well as the normalization parameters, this PR implements a class to address these issues.

Status

Ready

Types of changes

Non-breaking change (fix or new feature that would not break existing functionality).
Breaking change (fix or new feature that would cause existing functionality to change).
New tests added to cover the changes.
Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
Quick tests passed locally by running ./runtests.sh --quick --unittests.
In-line docstrings updated.
Documentation updated, tested make html command in the docs/ folder.

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv · 2021-07-16T09:57:17Z

The initial implementation requires joblib, let me choose another way and then reopen this PR.

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv · 2021-07-16T16:03:48Z

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here?

madil90 · 2021-07-19T02:24:33Z

/build

madil90 · 2021-07-19T08:00:48Z

/build

wyli · 2021-07-19T08:57:53Z

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here?

I think this could be part of monai/data/utils.

The implemention could use pytorch or monai dataloader instead of the multiprocessing lib
it's currently hard coded to use a datalist as in put, but it should be more flexible to take monai dataset as an input
would be great to reuse this utility to generate the relevant statistics of the dataset,

MONAI/monai/handlers/utils.py

Line 200 in f943ab4

{

instead of hardcoding the type of stats to be median/min/max

monai/data/dataset_calculator.py

yiheng-wang-nv · 2021-07-20T14:36:47Z

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here?

I think this could be part of monai/data/utils.

The implemention could use pytorch or monai dataloader instead of the multiprocessing lib

it's currently hard coded to use a datalist as in put, but it should be more flexible to take monai dataset as an input

would be great to reuse this utility to generate the relevant statistics of the dataset,

MONAI/monai/handlers/utils.py

Line 200 in f943ab4

{

instead of hardcoding the type of stats to be median/min/max

Thanks @wyli . My question is that when the dataset has images in different spatial sizes, it cannot support batch_size > 1 directly. Therefore, this PR used multiple processing instead. Do you have any suggestions?

wyli · 2021-07-20T14:39:02Z

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here?

I think this could be part of monai/data/utils.

The implemention could use pytorch or monai dataloader instead of the multiprocessing lib

it's currently hard coded to use a datalist as in put, but it should be more flexible to take monai dataset as an input

would be great to reuse this utility to generate the relevant statistics of the dataset,

MONAI/monai/handlers/utils.py

Line 200 in f943ab4

{

instead of hardcoding the type of stats to be median/min/max

Thanks @wyli . My question is that when the dataset has images in different spatial sizes, it cannot support batch_size > 1 directly. Therefore, this PR used multiple processing instead. Do you have any suggestions?

the dataloader is implemented with a proper multiprocessing support, I think you can use it with batch_size=1

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv · 2021-07-28T05:46:01Z

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here

Hi all, could you please help to review this PR?
I think this is needed, but I'm not sure if I put the class into the right place. Or should I use a tutorial rather than a MONAI feature here?

I think this could be part of monai/data/utils.

The implemention could use pytorch or monai dataloader instead of the multiprocessing lib

it's currently hard coded to use a datalist as in put, but it should be more flexible to take monai dataset as an input

would be great to reuse this utility to generate the relevant statistics of the dataset,

MONAI/monai/handlers/utils.py

Line 200 in f943ab4

{

instead of hardcoding the type of stats to be median/min/max

Hi @wyli , thanks for the comments. The code may not be able to be in monai/data/utils, since it needs to import dataset and dataloader and putting into monai/data/utils could lead to circular import.

Now I've switched to use monai dataloader.
The input has been changed to use monai dataset.
As for the write_metrics_reports handler, it needs to calculate with all accumulated image intensities. I think the way I updated in the latest PR may be faster (see calculate_statistics function with can get mean, std, max and min and do not need to accumulate all intensities).

Could you please help to re-review the code, thanks!

yiheng-wang-nv · 2021-07-28T05:46:58Z

Hi @ericspod , I've updated the code according to your comments, could you please help to re-review it, thanks!

Signed-off-by: Yiheng Wang <[email protected]>

wyli · 2021-07-28T14:17:25Z

looks like a useful utility, I'd suggest changing its name to DatasetSummary, what do you think?

ericspod · 2021-07-28T23:45:53Z

monai/data/dataset_calculator.py

+            all_intensities.append(intensities)
+
+        all_intensities = list(chain(*all_intensities))
+        self.data_min_percentile, self.data_max_percentile = np.percentile(


We're still accumulating all the foreground pixel values from all the images here, unless we have some way of doing an incremental percentile calculation we are stuck with that.

monai/data/dataset_calculator.py

ericspod

I think we're good with a few minor changes possibly.

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv · 2021-07-29T15:18:28Z

Hi @wyli , I've changed the name, and also add a test case for anisotropic spacings.

Signed-off-by: Yiheng Wang <[email protected]>

…nv/MONAI into 2583-add-dataset-calculator

wyli · 2021-07-30T08:58:06Z

Perhaps in another iteration we can change the code to use some online histogram estimation such as https://github.com/maki-nage/distogram

yiheng-wang-nv · 2021-07-30T14:10:58Z

Mark: later I will update a tutorial in https://github.com/Project-MONAI/tutorials/tree/master/modules/dynunet_pipeline to show how can we get the spacings and normalization parameters.

* Add DatasetCalculator Signed-off-by: Yiheng Wang <[email protected]> * update docstring Signed-off-by: Yiheng Wang <[email protected]> * use multiprocessing Signed-off-by: Yiheng Wang <[email protected]> * update to use dataset and other places Signed-off-by: Yiheng Wang <[email protected]> * update to support array return Signed-off-by: Yiheng Wang <[email protected]> * update with new testcases and change name Signed-off-by: Yiheng Wang <[email protected]> * update min test Signed-off-by: Yiheng Wang <[email protected]> * update unittest Signed-off-by: Yiheng Wang <[email protected]> * fix vstack error Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv added 2 commits July 16, 2021 17:15

Add DatasetCalculator

8f411d0

Signed-off-by: Yiheng Wang <[email protected]>

update docstring

1921f6f

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv closed this Jul 16, 2021

use multiprocessing

7ac7159

Signed-off-by: Yiheng Wang <[email protected]>

yiheng-wang-nv reopened this Jul 16, 2021

yiheng-wang-nv requested review from Nic-Ma, ericspod, rijobro and wyli July 16, 2021 15:36

ericspod reviewed Jul 19, 2021

View reviewed changes