Inference (2D & 3D)
The inference phase is where your SABER workflow pays off - applying trained models or zero-shot segmentation to generate high-quality 2D and 3D segmentations on new datasets. Whether you're using raw SAM2 capabilities or your custom-trained classifier, SABER provides flexible options for both semantic and instance segmentation.
🎯 Segmentation Modes
SABER offers two distinct approaches for inference, depending on whether you have a trained classifier:
Expert-Guided Segmentation
Apply your custom-trained classifier for domain-specific results:
- Semantic segmentation: Identify all learned biological classes simultaneously.
- Instance segmentation: Focus on specific objects of interest with individual instance masks.
Semantic vs Instance Segmentation
Semantic Segmentation (--target-class -1):
- Assigns every pixel to a biological class (background, lysosome, carbon, etc.)
- Produces a single mask with different colors/values for each class
- Ideal for understanding tissue composition and spatial relationships
- Answers: "What types of structures are present and where?"
Instance Segmentation (--target-class N where N > 0):
- Identifies individual objects of a specific class
- Each object gets a unique instance ID
- Enables counting, size analysis, and individual object tracking
- Answers: "How many lysosomes are there and what are their properties?"
🖼️ 2D Micrograph Segmentation
Basic Usage
Segment individual images or entire folders of micrographs:
# Single micrograph
saber segment micrographs \
--input path/to/image.mrc \
--output segmentation_results.zarr
# Batch processing
saber segment micrographs \
--input 'path/to/micrographs/*.mrc' \
--output batch_segmentations.zarr
Parameters:
| Parameter | Description | Example |
|---|---|---|
--input | Path to micrograph or folder with file extension | 'path/*.mrc' |
--output | Path to output Zarr file | results.zarr |
--model-config | Path to classifier model config | results/model_config.yaml |
--model-weights | Path to trained classifier weights | results/best_model.pth |
--target-class | Target class (-1 for semantic, N>0 for instance) | 2 |
--target-resolution | Target resolution in Angstroms | 10 |
🧊 3D Tomogram Segmentation
Basic Usage
Generate 3D segmentations using copick-managed data:
saber segment tomograms \
--config copick_config.json \
--model-config results/model_config.yaml \
--model-weights results/best_model.pth \
--segmentation-name organelles \
--target-class 2
When no --run-ids are provided, SABER will segment the entire project and save the results under provided --segmentation-name and --segmentation-session-id flags (by default, the user-id will always be saber. )
Parameters:
| Parameter | Description | Example |
|---|---|---|
--config | Path to Copick config file | copick_config.json |
--voxel-size | Resolution of tomograms to process | 10 |
--tomo-alg | Reconstruction algorithm to query | denoised |
--slab-thickness | Thickness of slab for initial segmentation | 10 |
--segmentation-name | Name of segmentation session | organelles |
--segmentation-session-id | Session ID for segmentation mask | 1 |
--model-config | Path to classifier model config | results/model_config.yaml |
--model-weights | Path to trained classifier weights | results/best_model.pth |
--target-class | Target class (-1 for semantic, N>0 for instance) | 2 |
--num-slabs | Number of slabs to segment | 1 |
--run-ids | Specific tomogram runs to process | None |
🎉 Congratulations! You now have the complete SABER workflow - from raw data to production-ready segmentations. For advanced usage and API integration, check out the Python API Documentation.