Pre-processing Your Data

Pre-processing is the first step in the SABER workflow, where you'll prepare your electron microscopy (EM) or cryo-electron tomography (cryo-ET) data for segmentation and annotation. This step leverages SAM2 (Segment Anything Model 2), a foundation model that can segment arbitrary structures without domain-specific training, making it ideal for scientific imaging applications.

🗂️ Supported File Types

SABER can read and process the following file formats:

MRC (.mrc)
TIFF (.tif, .tiff)
Zarr (.zarr)

Material science formats:

SER (.ser)
DM3/DM4 (.dm3, .dm4)
...and more!

🔍 Quick Assessment: Exploring Raw SAM2 Segmentations

Foundation Model Power: SAM2 is a breakthrough foundation model developed by Meta AI that can segment objects across diverse domains with remarkable zero-shot capabilities. Despite being trained primarily on natural images, SAM2 can identify and delineate structures in electron microscopy data without any domain-specific training.

Before committing to the full SABER workflow, quickly assess whether your data works well with SAM2's segmentation capabilities:

For Micrographs (2D Data)

For micrograph (2D data), we have the option to either directly segment the full image or downsample the data prior to running segmentation. *.mrc files or other file formats where the metadata is available we can provide the --target-resolution flag in Angstrom.

saber segment micrographs \
    --input path/to/image.mrc \
    --target-resolution 10 # Angstroms

For *.tif images, or data where the pixel size is unavailable in the meta-data, we need to provid the scale that we want to downsample our data.

saber segment micrographs \
    --input path/to/image.mrc \
    --scale 3 # Downsample by factor of 3

For Tomograms (3D Data)

SABER uses copick to access tomographic volumes. You have two options for assessment:

Single slab assessment (quick preview at a central z-depth):

saber segment slab \
    --config config.json \
    --voxel-size 10 --tomo-alg denoised --slab-thickness 10 \
    --run-id Position_10_Vol

Full tomogram segmentation (segment the entire 3D tomgoram):

saber segment tomograms \
    --config config.json \
    --voxel-size 10 --tomo-alg denoised --slab-thickness 10 \
    --run-id Position_10_Vol

SABER's tomographic workflow begins with 2D segmentation initialization, then propagates segments into 3D. The --slab-thickness flag (in voxels) averages density within a specified thickness, improving segmentation quality by reducing noise. If structures segment well in 2D slabs, they typically translate well to full 3D segmentation.

This preview helps you understand what structures SAM2 naturally identifies in your data and plan your annotation strategy accordingly.

🧬 Pre-processing Electron Micrographs

For single-particle datasets, ADF/BF signals from S/TEM, or FIB-SEM micrographs -- use the saber classifier prep2d command:

saber classifier prep2d \
    --input 'path/to/*.mrc' \
    --output training.zarr

Dense segmentation approach: This command performs comprehensive segmentation across all input images, identifying every discernible structure that SAM2 can detect. The result is a complete inventory of potential biological features ready for expert classification.

🧩 Generating Training Data with Initial SAM2 Segmentations

💡 Why SABER's Preprocessing is Different

Traditional workflows require you to manually draw every mask from scratch. SABER precomputes ALL possible segments using SAM2's foundation model, then lets you focus on the science -- simply assigning biological meaning to structures that are already perfectly segmented.

Generate comprehensive slab-based segmentations that maintain 3D context:

saber classifier prep3d \
    --config config.json \
    --zarr-path output_zarr_fname.zarr \
    --num-slabs 3

This will save multiple slab-wise segmentations across different z-depths, stored in a Zarr volume format that preserves both the original data and comprehensive SAM2 masks. This multi-depth approach ensures comprehensive coverage of all potential biological features for expert annotation.

SABER GUI

💡 Why multiple slabs?

Small objects or sparse structures might not be present in a single slab projection. By generating multiple 2D slab projections at different z-depths, SABER captures as many segmentations and instances of your target objects as possible. This is particularly important for:

Small organelles that appear sporadically through the volume
Thin structures that might be missed in thick slab averages
Objects with variable density that become more visible at certain depths

🎨 Annotation with the SABER GUI

Launch the GUI to begin annotating your pre-processed data:

saber gui --input output_zarr_fname.zarr

Once preprocessing is complete, SABER's unique annotation workflow begins. Instead of drawing masks from scratch, you simply:

How the GUI works:

Point and Click on the precomputed SAM2 segmentations.
Assign Class Labels using the menu on the right.
Save the Annotations Save the resulting JSON file with the bottom right button.

SABER GUI

Annotation Guidelines - How Many Images to Annotate?

We recommend 20-40 runs per dataset
Lower range (20): When multiple objects appear per image/slab
Higher range (40): When only few instances are available
Consistency is key: Maintain uniform criteria across all annotations
Handle ambiguous segments: When uncertain, prefer skipping over mislabeling

Tip: For transferring data between machines, it's recommended to compress your Zarr files:

zip -r curated_labels.zarr.zip curated_labels.zarr

🏷️ Applying Annotations for Classifier Training

Once you've completed annotations in the GUI, use the labeler command to apply your JSON annotations to the SAM2 masks, creating a training-ready dataset. The labeler converts your point-and-click annotations into properly indexed training data, handling class ordering automatically or according to your specifications.

Basic Usage

saber classifier labeler \
    --input training.zarr \
    --labels labels.json \
    --classes lysosome,carbon,edge \
    --output labeled.zarr

We can either control the ordering of the labels or apply a subset of the labels with the --classes flag. If the flag is omitted, all classes are used in alphabetical orde

Ready to move on? Check out the Training a Classifier tutorial!