This repository accompanies research from the Zaritsky Lab of Computational Cell Dynamics on interpretability and confidence estimation for in silico labeling predictions. It provides two complementary applications of Mask Interpreter: a supervised confidence model that estimates patch-level in silico labeling performance using the predicted fluorescence image and its corresponding importance mask, and a single-cell workflow that generates importance masks and explanation signatures for single-cell in silico labeling predictions.
Preprint link will be added upon publication.
The two applications address different questions and are provided as independent workflows, each with dedicated code, configuration, environment, and documentation.
The supervised confidence workflow estimates patch-level in silico labeling performance from two inputs: the in silico labeling prediction and its corresponding importance mask. A 3D regression model predicts the in silico labeling error, which is converted into a confidence score for identifying predictions that may be unsuitable for downstream biological analysis.
View the Supervised Confidence documentation
The Single-Cell Mask Interpreter workflow applies Mask Interpreter to single-cell in silico labeling predictions. It generates importance masks that identify image regions important for preserving each prediction, enabling explanation signatures to be examined at single-cell resolution.
View the Single-Cell Mask Interpreter documentation
MaskInterpreter-Applications/
├── README.md
├── LICENSE
├── images/
├── Supervised_Confidence/
└── Single_Cell_Mask_Interpreter/
The repository is organized into two independent workflows:
Supervised_Confidence/contains the notebooks, source code, configuration templates, and documentation for training and applying the supervised confidence model.Single_Cell_Mask_Interpreter/contains the notebooks, source code, configuration templates, and documentation for applying Mask Interpreter to single-cell in silico labeling predictions.
Large data files, pretrained models, and precomputed analysis results are distributed separately through Zenodo. Their download and configuration are described in the relevant workflow documentation.
Large microscopy data, pretrained model checkpoints, and precomputed results are distributed separately and are not tracked directly in this GitHub repository.
The original Mask Interpreter Zenodo record provides the shared resources used by these workflows, including pretrained in silico labeling and field-of-view Mask Interpreter models, Nuclear-envelope field-of-view example data, and train/test lists for the full datasets.
Resources introduced by the applications in this repository are available through the Mask Interpreter Applications Zenodo record:
Supervised_Confidence.zipcontains the pretrained supervised confidence models and precomputed result arrays.Single_Cell_Mask_Interpreter.zipcontains the single-cell example data and pretrained models required for the Single-Cell Mask Interpreter workflow.
The exact files required by each workflow and their expected directory configuration are described in the dedicated documentation:
Full microscopy datasets are not redistributed through this repository or its companion Zenodo record. Instructions for obtaining and preparing the full datasets from their original sources are provided in the relevant workflow documentation.
Clone the repository:
git clone https://github.com/zaritskylab/MaskInterpreter-Applications.git
cd MaskInterpreter-ApplicationsThen choose one of the two independent workflows below. Each workflow has its own environment, resource requirements, configuration, and demonstration instructions:
- Supervised Confidence: installation and workflow
- Single-Cell Mask Interpreter: installation and workflow
If you use this repository, please cite the associated paper and repository.
@misc{Trustworthy_in_silico_labeling_2026,
title = {Trustworthy in silico labeling via semantic visual interpretability of image-to-image translation},
author = {Miller, Gad and Ben Nedava, Lion and Zaritsky, Assaf},
year = {2026},
howpublished = {\url{https://github.com/zaritskylab/MaskInterpreter-Applications}},
doi = {10.5281/zenodo.20522083}
}@article{Trustworthy_in_silico_labeling_2026,
title = {TODO},
author = {TODO},
journal = {Preprint},
year = {2026},
doi = {TODO}
}This work was carried out in collaboration with Lion Ben Nedava. Related MaskInterpreter code can be found here:
https://github.com/zaritskylab/MaskInterpreter
The single-cell in silico labeling component builds on the CELTIC framework:
https://github.com/zaritskylab/CELTIC
For questions, please contact:
- Gad Miller: gadmiller7@gmail.com
- Lion Ben Nedava: lionben89@gmail.com
- Prof. Assaf Zaritsky: assafzar@gmail.com
This repository is intended for academic and research use and is licensed under CC BY-NC 4.0. See LICENSE for details.