Quick Start Guide

This guide walks you through a complete Saber workflow: from curation of training data to segmenting micrographs or tomograms in both 2D and 3D.

Basic Workflow Overview

The SABER workflow consists of three main phases:

1. Data Preparation & Annotation - Generate initial segmentations and curate training labels
2. Model Training - Train a domain expert classifier on your annotated data
3. Production Inference - Apply your trained model to new data

For reference, you can skip steps 1 and 2 to visualize the raw SAM2 segmentations in 2D or 3D without a domain expert classifier.

🧩 Phase 1: Curating Training Labels and Training and Domain Expert Classifier

Producing Intial SAM2 Segmentations

Use prep3 to generate 2D segmentations from a tomogram using SAM2-style slab-based inference. These masks act as a rough initialization for downstream curation and model training.

For tomogram data:

saber classifier prep3d \
    --config config.json \
    --voxel-size 10 --tomo-alg denoised \
    --num-slabs 3 --output training.zarr \

For electron micrograph/single-particle data:

saber classifier prep2d \
    --input path/to/folder/*.mrc \
    --ouput training.zarr \
    --target-resolution 10 

In the case of referencing MRC files from single particle datasets use prep2d instead.

🎨 Annotating Segmentations for the Classifier with the Interactive GUI

Launch an interactive labeling session to annotate the initial SAM2 segmentations and assign class labels.

saber gui --input training.zarr 

For transfering the data between machines, its recommended ziping (compressing) the zarr file prior to data transfer (e.g. zip -r training.zarr.zip training.zarr).

After you download the anntoated JSON file, you can apply the annotations on the original zarr file.

saber classifier labeler \
    --input training.zarr \
    --labels labels.json \
    --classes class1,class2,class3 \
    --output labeld.zarr

Once the labeled zarr is available, split the dataset into training and validation sets:

saber classifier split-data \
    --input labeled.zarr \
    --ratio 0.8

🧠 Phase 2: Train a Domain Expert Classifier

Train a classifier using your curated annotations. This model improves segmentation accuracy beyond zero-shot results by learning from expert-provided labels.

saber classifier train \
    --train labeled_train.zarr --validate labeled_val.zarr \
    --num-epochs 75 --num-classes 4 

🔍 Phase 3: Inference

🖼️ Producing 2D Segmentations with SABER

SABER operates in two modes depending on your input: interactive mode when processing a single image, and batch processing mode when you provide a file path pattern (like --input 'path/to/*.mrc') to process entire datasets automatically.

saber segment micrographs \
    --input path/to/image.mrc \
    --model-config results/model_config.yaml \
    --model-weights results/best_model.pth \
    --target-resolution 10 # Angstrom 
    # provide --scale 3 instead if you want to dowsample by 3

🧊 Producing 3D Segmentations with SABER

For tomograms, SABER enters interactive mode when you specify particular --run-ids, or batch processes the entire project when --run-ids is omitted.

saber segment tomograms \
    --config config.json
    --model-config results/model_config.yaml \
    --model-weights results/best_model.pth \
    --target-class 2 --run-ids Position_12_Vol

What's Next?

This workflow gives you a quick introduction to the saber segmentation pipeline. To learn more:

• Checkout the tutorial.
• Learn how to use the API.