MaskInterpreter | Trustworthy in silico labeling via semantic visual interpretability of image-to-image translation

Lion Ben Nedava^1*, Gad Miller^1*, Nitsan Elmalam¹, Mateheus Viana³, Jianxu Chen², Nathalie Gaudreault³, Sussane Rafelski³, Assaf Zaritsky¹

*Equal contribution

1. Institute for Interdisciplinary Computational Science, Stein Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Beer-Sheva 84105, Israel
2. Leibniz-Institut fur Analytische Wissenschaften - ISADS - e.V, Dortmund, Germany
3. Allen Institute for Cell Science, Seattle, WA, USA

1. Abstract

Cross-modality image translation promises to provide multiple layers of biological information from a single input, yet its practical application is stalled by a lack of interpretability and the inability to account for model imperfections. In silico labeling, the inference of organelle localization from label-free images, is a primary example where this black-box nature limits adoption. We present Mask Interpreter, a generalized method for semantic visual interpretability of image-to-image translation models. By uncovering organelle-specific "explanation signatures", we demonstrate that models leverage unique and reproducible biological patterns. Mask Interpreter outperforms traditional xAI approaches, identifies batch effects and localized prediction errors when ground-truth fluorescence is unavailable. Our supervised confidence modeling provides fine-grained reliability assessment at single-cell resolution, enabling the automated exclusion of artifacts from downstream analyses. By bridging the gap between computational inference and meaningful biological features, Mask Interpreter transforms in silico labeling into a rigorous, evidence-based instrument for scientific discovery across diverse biomedical imaging modalities.

Figure 1. Interpreting in silico labeling using Mask Interpreter. (A) Training of in silico labeling models using matched label-free and fluorescence images. (B) Example of predictions with/without using MaskInterpreter’s importance mask. (C-F) Training and inference pipeline schematic.

See Paper (link) for details: https://doi.org/10.64898/2026.08.13.744455

2. Overview

Deep learning models often operate as "black boxes," making it difficult to understand which input features drive their predictions. MaskInterpreter addresses this by learning a per-organelle mask generator network that identifies important regions through a novel training objective:

• Preserve predictions: Important regions (high mask values) should be sufficient to maintain the model's original prediction
• Minimize mask size: The mask should be as minimal as possible, highlighting only truly essential regions
• Target correlation: Predictions on masked inputs should maintain a specified correlation with the original predictions

Key Features

• Self-supervised training - No ground truth explanations needed
• Model-agnostic - Works with any differentiable predictor (e.g. classifiers, regressors, image-to-image models)
• New measurement to quantify explanations - the Pearson correlation coefficient (PCC) between the predictions derived from the unperturbed input, and the predictions derived from the importance mask-induced noisy inputs

3. Repository Structure

mask_interpreter/
├── README.md
├── pyproject.toml              # Package dependencies and configuration
├── example.ipynb               # Quick start tutorial notebook
│
├── md/                         # Documentation files
│   ├── quickstart.md           # Quick start guide
│   ├── data.md                 # Full data download instructions
│   ├── usage.md                # Usage examples and training
│   ├── examples.md             # Additional examples
│   └── reproduce.md            # Reproducing paper figures
│
├── models/                     # Core MaskInterpreter implementations
│   ├── MaskInterpreter.py      # Image-to-image models
│   ├── MaskInterpreterRegression.py   # Regression models
│   ├── MaskInterpreterCLF.py   # Classification models
│   ├── regressor_cellcycle.py  # Cell cycle marker regression
│   ├── clf-cifar10.py          # CIFAR-10 classifier
│   └── UNETO.py                # U-Net architecture for mask generation
│
├── create_data/                # Data download and preparation scripts
│   ├── download_and_create_dataset_full.py
│   ├── download_and_create_dataset_singlecell.py
│   ├── create_metadata.py
│   └── segment_and_create_pertrub_dataset.py
│
├── figures/                    # Scripts to reproduce paper figures
│   ├── 0_reproduce_unet_scores.py
│   ├── 1_choose_noise_scale.py
│   ├── 2_choose_th.py
│   ├── 3_calculate_unet_scores.py
│   ├── 4_calculate_explanation_mask_efficacy.py
│   └── ...
│
├── gui/                        # Graphical user interface
│   ├── gui.py
│   └── gui_logic.py
│
├── utils/                      # Utility modules
│   ├── callbacks.py            # Training callbacks
│   ├── metrics.py              # Evaluation metrics (PCC, etc.)
│   └── utils.py                # Helper functions
│
├── dataset.py                  # Data loading utilities
├── global_vars.py              # Global configuration and paths
├── mg_analyzer.py              # Mask generator analyzer
├── train.py                    # Main training script
└── test.py                     # Testing utilities

4. General Installation and setup

Prerequisites

• Python 3.9+
• CUDA-compatible GPU (recommended)
• Conda (recommended for environment management)

Download Trained Models and Example Data

Pre-trained models and example data are available from Zenodo for quick start and reproducibility.

Download from Zenodo

Visit the Zenodo repository to download the required files:

Zenodo Link: https://zenodo.org/records/22107794

The archive contains:

• Pre-trained in silico labeling models - Trained on various organelles
• Pre-trained MaskInterpreter models - Corresponding interpretation models for each predictor
• Example data - Sample images with metadata for testing and validation
• Train and test lists of the full data

Extraction and Setup

After downloading, extract the contents and set the paths accordingly:

# Download the archive from Zenodo
wget https://zenodo.org/records/18590674/files/models_and_data.zip

# Extract to your desired location
unzip models_and_data.zip -d /path/to/your/directory

# The extracted files under models_and_data dir will contain:
# - example_data (example dataset with train/test CSV files)
# - models/ (pre-trained models)
# - train_test_list/ (train and test lists of the full data)

Make sure to update your environment variables (see next section) to point to these directories.

Setup

1. Clone the repository

git clone https://github.com/lionben89/cell_generator.git
cd cell_generator

2. Create conda environment

conda create -n maskinterpreter python=3.9 tensorflow-gpu=2.6
conda activate maskinterpreter

3. Install the package in editable mode

pip install -e .

This will install all dependencies from pyproject.toml automatically

Verify Installation

Note: Run this verification on the terminal of the computer/node with GPU access.

# Activate the environment
conda activate maskinterpreter

# Run verification script
python -c "
import warnings
warnings.filterwarnings('ignore')

import tensorflow as tf
print('GPUs Available:', tf.config.list_physical_devices('GPU'))

from models.MaskInterpreter import MaskInterpreter
print('MaskInterpreter imported successfully!')
"

Configuration

The project uses global_vars.py for configuration including paths for data, models, and the repository.

Open global_vars.py and update the path variables at the top of the file:

# ============================================
# Path Configuration
# ============================================
# Base paths - update these to match your environment
BASE_PATH = '/path/to/your/data'
DATA_PATH = '/path/to/your/data/train_test_list'
CWD = 'current working dir'

Note: Those variables must be set before importing any project modules.

5. Quick Start

6. Full Data Download

7. Usage

8. Additional Examples

9. Recreate Figures from Paper

PyTorch Implementation

PyTorch Implementation and supervised confidence model implementation: For a PyTorch version of MaskInterpreter and tools for assessing the supervised prediction quality at inference time using MaskInterpreter, see the companion repository: https://github.com/zaritskylab/Interpretability

Citation

If you use MaskInterpreter in your research, please cite:

@article {Ben Nedava2026.08.13.744455,
	author = {Ben Nedava, Lion and Miller, Gad and Elmalam, Nitsan and Viana, Matheus P. and Chen, Jianxu and Gaudreault, Nathalie and Rafelski, Susanne M. and Zaritsky, Assaf},
	title = {Trustworthy in silico labeling via semantic visual interpretability of image-to-image translation},
	year = {2026},
	doi = {10.64898/2026.08.13.744455},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2026/08/15/2026.08.13.744455},
	eprint = {https://www.biorxiv.org/content/early/2026/08/15/2026.08.13.744455.full.pdf},
	journal = {bioRxiv}
}

link:https://doi.org/10.64898/2026.08.13.744455

Acknowledgments

• Allen Institute for Cell Science for the cell imaging datasets and review of the paper.

Contact

• Email: assafzar@gmail.com , lionben89@gmail.com, gadmicha@post.bgu.ac.il
• Lab: Zaritsky Lab

License

This repository (data, documentation, and figures) is intended for academic and research use, and is licensed under CC BY-NC 4.0. See LICENSE for details.