From NITRC Wiki

Contents 1 JIST Use Cases 1.1 Introduction 1.2 How to Run 1.3 Image Segmentation 1.3.1 Geodesic Active Contour 2D 1.3.2 Geodesic Active Contour Without Edges 2D 1.3.2.1 Chan-Vese Model 1.3.2.2 Mumford-Shah Model 1.3.2.3 Rodin-Osher-Fatemi Model 1.3.2.4 Chan-Esedoglu-Nikolova Model 1.4 Pipeline Logic

JIST Use Cases

Introduction

• These use cases are examples of working applications of JIST. They are supported and maintained by Blake Lucas (blake@cs.jhu.edu). Please direct comments and suggestions towards him. For bug reports and feature requests, please use the Tracker.
• Use cases are only guaranteed to run with the head release of JIST's public plug-ins. They may or may not work for stable releases.
• Data is pulled from the web at runtime, so you'll need an internet connection in order to run these pipelines.
• All references are embedded in the Algorithm Information associated with each module.
• Extension and enhancement of modules is encouraged, but please do not commit unpublished algorithms to the public plug-in tree.

How to Run

1. Start the Pipeline Layout tool from MIPAV or Eclipse.
2. Open the layout file for the use case.
3. Go to Project->Layout Preferences and change the output path to a local directory.
4. Go to File->Save.
5. Go to Project->Process Manager to run the use case.
6. Select all experiments and press play.

Image Segmentation

Geodesic Active Contour 2D

• Segments an image using information about the likelihood of a pixel being part of the desired segmented region (pressure force) and edge information derived from image gradients (advection force).
• The layout for this use case can be found here.
• The pipeline will do the following:

1. Classify image intensities using a fuzzy threshold.
2. Calculate the gradient magnitude of the image.
3. Extend the gradient field using Gradient Vector Flow (vector field diffusion).
4. Initialize the segmentation with a circle.
5. Segments the image using a Geodesic Active Contour (GAC).
6. Visualizes the segmentation as a contour overlaid on the original image, vector field, and pressure force.

Geodesic Active Contour Without Edges 2D

• Segments an image by dynamically adjusting the appearance and shape of a segmentation until the estimated image best matches the original image.
• There are several classic models for modeling regional image appearance, which include Chan-Vese, Mumford-Shah, and Rudin-Osher-Fatemi models.

Chan-Vese Model

• For the Chan-Vese model, the foreground and background are assumed to have constant intensity.
• The layout for this use case can be found here.
• The pipeline will do the following:

1. Initialize the segmentation with a circle.
2. Estimate the appearance with either the mean or median statistic.
3. Evolve boundary using Geodesic Active Contour and repeat from previous until convergence.
4. Visualizes the segmentation as a contour overlaid on the original image and estimated image.

Mumford-Shah Model

• For the Mumford-Shah model, there is an L2 penalty on the model's deviation from the image and a L2 penalty on the gradient magnitude.
• The layout for this use case can be found here.
• The pipeline will do the following:

1. Initialize the segmentation with a circle.
2. Estimate the appearance with the Mumford-Shah (MS) denoising model.
3. Evolve boundary using Geodesic Active Contour and repeat from previous until convergence.
4. Visualizes the segmentation as a contour overlaid on the original image and estimated image.

Rodin-Osher-Fatemi Model

• For the Rodin-Osher-Fatemi model, there is a L2 penalty on the model's deviation from the image and a penalty on the gradient magnitude. The gradient magnitude penalty is also referred to as Total Variation (TV) minimization.
• The layout for this use case can be found here.
• The pipeline will do the following:

1. Initialize the segmentation with a circle.
2. Estimate the appearance with the Rodin-Osher-Fatemi (ROF) denoising model.
3. Evolve boundary using Geodesic Active Contour and repeat from previous until convergence.
4. Visualizes the segmentation as a contour overlaid on the original image and estimated image.

Chan-Esedoglu-Nikolova Model

• For the Chan-Esedoglu-Nikolova model, there is a L1 penalty on the model's deviation from the image and a penalty on the gradient magnitude. The L1 penalty is approximated using a convex regularization so that the energy function can be easily optimized.
• The layout for this use case can be found here.
• The pipeline will do the following:

1. Initialize the segmentation with a circle.
2. Estimate the appearance with the Chan-Esedoglu-Nikolova (CEN) denoising model.
3. Evolve boundary using Geodesic Active Contour and repeat from previous until convergence.
4. Visualizes the segmentation as a contour overlaid on the original image and estimated image.

Pipeline Logic

• Demonstrates basic logic operations that are available in JIST. Logical operations appear as yellow modules because they are executed in the same process as the process manager for efficiency.
• The layout for this use case can be found here.
• In the first experiment, the "Less Than" operation will return true, so the "Add" operation will be performed and the "Subtract" operation will be ignored. Because the subtract operation is not executed, The "Volume Join" module cannot be executed.
• In the second experiment, the "Less Than" operation will return false, so the "Subtract" operation will be performed and the "Add" operation will be ignored. In this case, the "Volume Join" will be executed since both its parents are executed.