Skip to content

Version 3.7.3

Latest

Choose a tag to compare

@danifranco danifranco released this 04 Nov 10:50
· 37 commits to master since this release

Version 3.7.3 Release Notes

Major Changes

Tracking

  • Added cell tracking to the instance segmentation workflow (PROBLEM.TYPE: INSTANCE_SEG with PROBLEM.TRACKING.ENABLE: True), in 2D and 3D. DATA.TEST.PATH is one time-lapse sequence (one image per frame, in name order). The model creates the instances of each frame, which are then linked across frames
  • Instances are linked with Trackastra's transformer, used as pretrained weights without the trackastra package (PROBLEM.TRACKING.LINKER_MODEL):
    • "trackastra_general_2d": 2D, trained on diverse live-cell data
    • "trackastra_ctc": 2D and 3D, trained on the Cell Tracking Challenge datasets
  • PROBLEM.TRACKING.MASKS_PATH tracks instance masks you already have, with no segmentation model
  • PROBLEM.TRACKING.ALLOW_DIVISIONS controls whether a cell can be linked to two cells in the next frame
  • Tracks are saved in PATHS.RESULT_DIR.TRACKING in the Cell Tracking Challenge result format (mask<t>.tif + res_track.txt) and as GEFF (tracks.geff)
  • Tracking evaluation: when PROBLEM.TRACKING.GT_PATH points to CTC ground truth (e.g. a 01_GT/TRA folder), the CTC DET and TRA metrics are computed, along with link and division precision, recall and F1
  • New templates: 2d_tracking.yaml and 3d_tracking.yaml
  • The wizard (BiaPy-GUI and Fiji plugin) now offers tracking for instance segmentation and checks that the test folder can be read as the frames of a time-lapse

STU-Net deep supervision

  • Added MODEL.STUNET.DEEP_SUPERVISION (as in nnU-Net): during training, the lower-resolution decoder stages also predict the output, and the loss is a weighted mean of all of them. Inference is not affected. Available for SEMANTIC_SEG, INSTANCE_SEG (except with the E_offset channel) and DETECTION

Minor Updates

  • The wizard and BiaPy share one implementation of the data checks (file listing, image consistency, instance output channels...) in biapy.data, so the front-ends check data exactly as BiaPy does when it loads it
  • PROBLEM.IMAGE_TO_IMAGE.MULTIPLE_RAW_ONE_TARGET_LOADER with DATA.VAL.FROM_TRAIN now gives a clear configuration error (the validation data must be given in DATA.VAL.PATH and DATA.VAL.GT_PATH)
  • The conda-forge recipe now matches the pyproject.toml dependencies (pydantic<2.14, pywavelets, einops, torch-fidelity)
  • Fixed docstrings across the code (swt_loss, data generators, stunet, rdbm...)
  • Updated AGENTS.md: wizard, citations and new tests, by @loreoregi (#175)
  • Added tracking tests (Test35/Test36 and test_tracking.py), data-check tests, and an API check for loading BMZ instance segmentation models by nickname

Bug Fixes

  • BiaPy.load_workflow_from_model("<bmz id>") and configurations with MODEL.SOURCE: bmz failed for instance segmentation models: the configuration was checked against the default channels (B+C) before the model's own channels were applied. BMZ instance segmentation models that don't declare their channels can now be used too
  • BMZ: the compatibility check couldn't find the model nickname in raw RDFs (as the zoo CI passes them), so the dependency check was skipped

Full Changelog: v3.7.2...v3.7.3