[Jiaqi-Lv]

Summary

This pull request adds the SegFormer-B5 tissue segmentation models distilled from Virchow2, plus the inference and export fixes those models need.

The models come from Computationally Efficient Pathology Segmentation using Knowledge Distillation from Foundation Models.

New model configs

pretrained_model.yaml registers five checkpoints from TIACentre/Virchow2_Distilled_Segmentation_Models:

Model	Dataset	Classes	Resolution	Patch
segformerb5_distilled_PUMA_1/2/5	PUMA	6 (background, stroma, blood vessels, tumour, epidermis, necrosis)	1.14 mpp	224, stride 112
segformerb5_distilled_IGNITE_1	IGNITE	7 (unknown, background, tumour, stroma, necrosis, macrophages, rest)	0.5 mpp	448, stride 224
segformerb5_distilled_BEETLE_1	BEETLE	5 (unknown, other, non-invasive epithelium, invasive epithelium, necrosis)	1.0 mpp	448, stride 224

SegFormer inference

• Segformer takes an optional class_dict.
• Forward no longer relies on the Hugging Face segmentation head's hooks. Those hooks are shared across DataParallel replicas and feed the same feature map into every decoder projection. The model now runs each encoder stage directly and passes those hidden states to decode_head.
• load_state_dict unwraps checkpoints nested under a model key.
• postproc accepts a Dask array, runs the morphology on a computed array, and returns a Dask array so the engine can write the mask to Zarr.

QuPath export

Semantic-segmentation GeoJSON now uses contour hierarchy, so a nested class becomes a hole in the surrounding polygon. Classes with no pixels are skipped.

Example

Notebook 06-semantic-segmentation runs segformerb5_distilled_PUMA_1 on a whole-slide image, writes an annotation store, and shows the overlay in TIAViz.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.91%. Comparing base (c0bbb14) to head (c15d4fe).

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #1131   +/-   ##
========================================
  Coverage    99.91%   99.91%           
========================================
  Files          112      112           
  Lines        13454    13470   +16     
  Branches      1610     1614    +4     
========================================
+ Hits         13443    13459   +16     
  Misses           5        5           
  Partials         6        6           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB