Image loading

[siemdejong]

fixes #3 #20

This PR changes the way the framework loads images. Images are loaded as part of the algorithm. The image loading strategy can be shared across multiple algorithms, but algorithms are not required to use it. An algorithm should adhere to the BaseSegmenter Protocol, but algorithms can use a predefined abstract base class if they require a numpy array (NumpySegmenter) or simply a path (PathSegmenter).

NumpySegmenters need to be initialized with mpp (among other things). This loads the image at that exact mpp. This relies on an unreleased PR to MONAI. The exact current commit sha that currently implements it is added to the dependencies. Ideally, this should be addressed later once Project-MONAI/MONAI#7574 is merged. Choosing specific mpp allows to better finetune the ideal hyperparameters for the algorithms and to circumvent the issue where images can different resolutions at level 0. Of course users can pick an mpp that is already in the pyramid, then no interpolation will be performed.

Segmenter.segment() requires to input a Path and output a geojson.FeatureCollection. This allows for multiple things:

1. an easy user interface; you can always input one thing and can expect it to work (unless the parameters are suboptimal)
2. the output is always the same which is up to the user how to use it
3. a fair comparison between runtimes

Any issues with the environment are now fixed, albeit somewhat unsafe via

[tool.uv]
index-strategy = "unsafe-best-match"

uv sync --all-extras is able to solve the environment with this change. Especially useful for the large_image dependency of histomicstk.

[siemdejong]

Below one single run of examples/unsupervised_benchmarking.py with an example tiff as explained in #19, so no CPGSegmenter result. With the changes introduced in this PR in place we can later tune the hyperparameters if necessary.

bg_subtractor_mog2_h500_v16_heatmap

entropy_masker_heatmap

fastsam_fastsam-x_heatmap

fesi_heatmap

grandqc_tissue_detection_mpp10_heatmap

hest_deeplabv3_default_mpp5_heatmap

histomicstk_saliency_heatmap

improved_fesi_heatmap

li_heatmap

morphological_disk3_heatmap

od_gmm_slide_heatmap

otsu_heatmap

watershed_mindist10_heatmap

yen_heatmap

[siemdejong]

I probably broke some other examples other than unsupervised_benchmark.py. They'd need to migrate to the new image loading requirements. If you want we can do that before merging.

[siemdejong]

I validated the workflow of adding a new algorithm. With this setup requires only a few lines and some import statements here and there.

https://github.com/RTLucassen/slidesegmenter

The model itself has a shift issue though, I don't think there's something we can do about it.

EDIT: well, maybe we can do something about it. It looks like it is more so a scaling issue. But haven't seen it with other algorithms. I can open an issue for that later.

[siemdejong]

EDIT: well, maybe we can do something about it. It looks like it is more so a scaling issue. But haven't seen it with other algorithms. I can open an issue for that later.

this actually highlights a bug. With the self.reader.get_wsi_at_mpp() it gets an image at a specific mpp, unless a relative tolerance allowes an existing level. Then, mask_to_geojson(mask, self.min_area, scaling_factor=self.reader.get_mpp(wsi, 0)[0] / self.mpp) still uses the self.mpp that was requests, which is misaligned. Will fix.

[siemdejong]

Fixed the bug by calculating the scaling factor using the model output. Ready to merge.

rtlucassen_slidesegmenter_heatmap