Prediction package for reproducible U-Net models, trained for semantic segmentation of microscopy images of root tissue from A. thaliana (https://github.com/qbic-pipelines/root-tissue-segmentation-core/). These models are trained using the mlf-core framework and tested for reproducibility. This package can be deployed within an analysis pipeline as a module for root tissue segmentation (rts) of fluorescence microscopy images.
This prediction module implements the Monte Carlo Dropout procedure (https://arxiv.org/abs/1506.02142) to calculate prediction uncertainty (uncertainty maps). Additionally, this module uses the Guided Grad-CAM algorithm (https://arxiv.org/abs/1610.02391) to compute input feature importance visualizations (interpretability maps), as implemented by the Captum library (https://captum.ai/).
- This package can be installed via pip: https://pypi.org/project/root-tissue-seg-package/
- The trained pytorch model used for prediction can be found here: https://zenodo.org/record/6937290/
- Segmentation prediction CLI:
rts-pred - Uncertainty of prediction CLI:
rts-pred-uncert - Input feature importance (Guided Grad-CAM) CLI:
rts-feat-imp
rts-pred -i ./brightfields -o ./predictions -m mark1-PHDFM-u2net-model.ckpt --suffix ""rts-pred-uncert -i ./brightfields -o ./predictions -m mark1-PHDFM-u2net-model.ckpt --suffix "" -t 5rts-feat-imp -i ./brightfields -o ./predictions -m mark1-PHDFM-u2net-model.ckpt --suffix "" -t 2
- Free software under MIT license
- Documentation: https://rts-package.readthedocs.io.
This package was created with mlf-core using cookiecutter.