Mask Interpreter models and example data

Trustworthy in silico labeling via semantic visual 1 interpretability of image-to-image translation (link to paper)

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 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 original predictions

Key Features

• Self-supervised training - No ground truth explanations needed
• Model-agnostic - Works with any differentiable predictor (classifiers, regressors, image-to-image models)
• Quantifiable explanations - Mask efficacy measured via Pearson Correlation Coefficient (PCC)

The Repo includes

1. Trained models
   1. mg_*  files are trained mask interpreter models , per organelle.
   2. unet_* files are trained in silico labeling models, per organelle.
2. Example data
3. Train and test lists of the full data