dyollb / segmantic Goto Github PK

Use the new bundle approach (example) to parametrize transforms using json/yml files.

• implement loading from json or yml
• post-processing (invert input transforms)?

https://github.com/microsoft/pyright/blob/main/docs/configuration.md#sample-pyprojecttoml-file

As a user I would like to train a network to detect landmarks in medical images.

• e.g. fiducials in head MRI to position 10-10 system

Checkout gh action of osparc-simcore-python-client

notice:

• using actions/setup-python@v2
• targets multiple python version strategy.matrix
• uses cache
  It is actually a basic extension of github actions tutotirla for python

Originally posted by @pcrespov in #3 (comment)

Same for

Currently the mean Dice and confusion matrix is computed for each image separately. As a user it would be nice to get

• aggregate statistics, e.g. average, min, max, or histograms with standard deviation (#38)
• additional metrics (e.g. Hausdorff)
• make configurable, e.g. add options to enable/disable each metric

setting data["dataset"] = str(dataset_path) is not overriding/overwriting keys image_dir and labels_dir. The training then uses those instead of the k-fold dataset.

how to fix:

data.pop("image_dir")
data.pop("labels_dir")

Why to switch

• ITK is slow to load and provides poor auto-completion in IDEs (e.g. VS Code), making it difficult to use.
• Most required functionality is covered by SimpleITK

Use case:

• load t1, t2 -> concatenate to image in pre-processing transforms (LoadImaged(keys=["t1", "t2", "label"], ConcatItemsd(...), )

Add light weight test data to releases or use generated data (e.g. like this)