SegBio

SegBio is a complete end-to-end toolkit for segmenting and separating long, narrow, bendy objects like C. elegans worms

The Challenge: Slow Annotation

Instance segmentation for biomedical images is powerful, but creating the training data is a major bottleneck. Manually tracing every single object in a clustered image is time-consuming, tedious, and scales poorly.

Our Solution: An Efficient Hybrid Pipeline

This pipeline is designed to maximize efficiency by combining rapid, inexact user input with powerful post-processing.

The workflow is broken into five key stages:

1. Rapid Annotation: Instead of fully tracing each worm, the user provides partial annotations (midline and width) using a custom GUI (AnnotationGUI.exe).
2. Data Generation: The simple annotations are extrapolated into a full training masks.
3. Model Training (PyTorch): A highly configurable U-Net (FlexiUnet.py) is trained to predict three separate channels from a single input image. The network's architecture is flexible, allowing for easy adjustment of its depth and filter count to scale the model's capacity. The model is trained to output: Foreground Mask: The body of the worms. Boundary Map: The pixels separating touching worms. Seed Map: A unique "center" or skeleton for each worm.
4. Inference & Post-processing: The 3-channel output from the U-Net is fed into a watershed algorithm. The seeds mark the start of each "basin," and the boundaries act as dams allowing the watershed to robustly separate instances that are touching in the foreground mask.
5. Human-in-the-Loop (Napari): A final standalone GUI (worm_editor_gui) loads the inference results and allows the user to quickly correct any errors (e.g., merging or splitting instances).

Key Features

• Efficient Annotation Strategy: Solves the "slow labeling" problem with a partial-to-full data generation pipeline.
• Robust Instance Separation: Uses the classic (Foreground, Boundary, Seed) + Watershed technique to reliably separate clustered instances.
• Flexible Model: The 'FlexiUnet.py' is parameterized for variable depth ('--depth') and filter count ('--base-filters'), making it easy to reconfigure.
• Model Training: Includes augmentations ('segmentor_utils.py'), a multi-channel loss function, validation splits, and checkpointing ('TrainFlexiUnet.py').
• Automated Post-processing: Includes an inference-time step ('postproc.py') that automatically filters and cleans the raw model output by removing small, noisy detections, and refining boundaries.
• Standalone Inference GUI: Provides a simple self-contained Python-based app that doesn't require any dependencies or even a python installation. Allows end-users to load their own images, run the full segmentation pipeline, and get immediate results.

Quickstart

• Download the standalone app from here
• Keep the weights file (.pth) in the same folder as WormGUIApp.
• Run WormGUIApp.exe

Tech Stack

• Python
• PyTorch
• scikit-image & scipy
• NumPy
• OpenCV
• Napari
• MATLAB & MATLAB App Designer (legacy annotation GUI)