hip-satomi / microbeseg Goto Github PK

Enables inference on servers without the need for a connection with graphical user interface.

Required:

• omero settings and password prompt
• omero project/dataset/file ids (if project/dataset id is given, process all files in the project/dataset)
• variables for model path, model, result path, overwriting, upload
• option to process local .tif files with specified format instead of using an OMERO connection

Window logo not loaded in message boxes.

Only for the owner "true" is returned.

Enables training on servers without the need for a connection with a graphical user interface.

So far, microbeSEG extracts the selected channel from the images. However, there may be cases, where the rgb information should be kept for annotation and model training, e.g. for histopathological images.

Goal: branch with rgb support (data cropping, model training + augmentation, inference)

Path.cwd() is not a good choice since it may return another directory than the microbeSEG directory.

Upload on zenodo models trained on hip-satomi, omnipose and cell tracking challenge data.

Adjust readme to cite also omnipose and cell tracking challenge when model trained on omnipose is used.

There is a normalization problem for meandering cell shapes:

https://forum.image.sc/t/omero-py-group-switching/65162/3

• for faster upload times, polygons are uploaded in a single plane. The OMERO.insight client cannot yet view correctly multiple ROIs within a single plane
• the OMERO.web client works fine and is therefore recommended

When copying images in OMERO, e.g., into another dataset/project, the copied image is just a link to the original image. In that case, 'getProject' returns a NoneType Object without attribute 'getName'. Use 'listParents' to see if an image has been copied and is linked to multiple datasets/projects.

For some data upsampling may be usefull for annotation and segmentation, e.g., object width smaller than 3-4 px. In addition, some data just differ in resolution and down- or upsampling may enable to use another available already trained model without the need for new annotations.

Some menus do not adapt well when maximizing them.

no window maximization icon on windows but a useless "?" icon

Could be benefitial to retrain one of the published models instead of training from scratch. This needs some deep evaluation (learning rates of the first deep layers vs output layers, layer freezing, ...).