Interactive Segmentation Using the Insight Segmentation Editor Tool

This document introduces the Segmentation Editor (SE) tool, an interactive application developed in Itk and VTK for semi-automated segmentation of image data.

The SE tool is a special editor which allows you to construct a labeled image using the output from the Insight watershed segmentation algorithm. Hand labeling of images is also supported. The goal of the software is to combine human interaction with automated image processing techniques to produce a better result than either technique can give on its own.

All of the image processing in the SE tool is done using Insight filters. Visualization is handled by VTK. The GUI was constructed with Tk

The source can be found in the InsightApplications/SegmentationEditor directory of the Insight distribution. The file "InsightApplications/SegmentationEditor/README.txt" describes how to build and run the application.

Watershed Segmentation
A brief introduction to watershed segmentation is useful in understanding the software and its purpose.

Watershed segmentation gets its name from the manner in which the algorithm segments regions of an image or volume into "catchment basins". These basins are low points in the intensity of the image being segmented. The basins are referred to as "segments" or "regions". These regions share boundaries with one another.

Imagine water raining into these basins. As the water level rises, the basins fill and water spills across the boundaries, joining those basins into larger basins. In this way, we can view the segmentation of the image into basins at different levels of scale, a continuous measure specified by the level of the water.

The output of the watershed segmentation technique is then a hierarchy of basins which can be viewed at different levels of scale. What remains is for a human to determine which basins at which scale levels represent the structure or structures of interest. The SE application is designed to facilitate that process and allow the user to piece together a binary mask of the structure from the hierarchy of segmented basins.

The input to the watershed algorithm is critical to the quality of its output. The watershed algorithm expects a height image as input. A "height image" is defined in this context as an image where higher image intensities correspond to logical boundaries between regions of interest. The algorithm does not presume to construct this input image for you because it cannot know what image features are interesting for your application. It is your job, therefore, to construct an input based on image features of interest.

The SE application leads you through the process of constructing a height image based on gradient magnitude which can be used as input to the segmentation algorithm. Gradient magnitude calculations on an image are commonly used in image processing applications to find object edges. Areas of high gradient magnitude are areas of rapid intensity change in the image, where objects are often set off from their background in relatively high relief. The smoothing filters used in the preprocessing stages of the SE application are tuned to preserve areas of high gradient magnitude which may be of interest.

A Semi-Automated Segmentation Method
The watershed segmentation does not target any particular region of an image; it processes the entire image. In order to label a particular structure in an image, the output of the watershed segmentation must be manipulated by hand to target voxels of interest. Because of the necessary human interaction, we refer to this process as semi-automated.

The SE tool is a test of the hypothesis that through image analysis and targeted visualization of the data, the task of hand labeling voxels contained in anatomical structures can be more efficient than hand segmentation alone. The SE tool is an implementation of the semi-automated method for watershed segmentation.

Segmentation with the SE Software
The SE software interface is organized into several modules which separate the preprocessing, segmenting, and editing stages of the segmentation process. The output from one module becomes the input to the next. At the end of this pipeline, you can use the editor to create labeled images. Tutorial-style documentation is provided within the interface itself.

Some screen captures of the SE software are given below. The data shown is taken from the Harvard Brigham and Women's Hospital Brain Tumor Segmentation Database.

Editor module
the Preprocessor module

Preprocessor module
the Editor module

Origins of the SE Software
This tool was developed at the University of Utah's Scientific Computing and Imaging Institute and was used as part of a validation study of the Insight implementation of the watershed algorithm. A full account of the validation study and its results will be available with the October 2002 Insight Version 1.0 release.

For more information on watershed segmentation and anisotropic diffusion, see the related Applications web pages in this section.

Much more detailed information about the filters used in the SE application is available in the Insight Manual Pages, which can be found under the Documentation section of

Go to next application.