ABSTRACT
Brain tissue segmentation is important in many medical image applications. We augmented the Expectation-Maximization segmentation algorithm in Slicer3 (www.spl.harvard.edu) . Currently, in the EM Segmenter module in Slicer3 user input is necessary to set tissue-class (Gray Matter, White Matter etc.) intensity values. Our contribution to the current pipeline is to automatically compute such values using k-means clustering. Our method can be applied to scans of varying intensity profiles and thus we obviate the need for a normalization step. We applied this pipeline on multiple datasets and our method was able to accurately classify tissue-classes. The implementation was done as a standalone utility in the Python programming language (www.python.org) and work is underway to incorporate it in the EM processing pipeline. 

EM Segmentation: Automatic Tissue Class Intensity Initialization Using K-means

Ultrasound imaging is the most popular and safe diagnostic medical imaging modality. Over the past years, a considerable number of studies have been conducted on new medical ultrasound imaging modalities such as ultrasound-based elasticity imaging, photoacoustic imaging, thermal imaging, etc. Moreover, results of such advanced ultrasound research are readily integrated with Image Guided Therapy (IGT) systems due to the advantages of ultrasound imaging such as mobility and real-time data acquisition. For integrating two systems, a communication method between systems has to be devised. OpenIGTLink is a standardized TCP/IP-based protocol for the integration of medical imaging and IGT systems. However, OpenIGTLink is not sufficient for communicating between IGT and advanced ultrasound research systems, because OpenIGTLink does not define any message type for general ultrasound data and some results of ultrasound researching.
In this paper, we propose an extension to OpenIGTLink, called OpenIGTLinkMUSiiC, by adding new message types and new network classes. Our new message types are designed for containing raw ultrasound data, advanced ultrasound modalities, and text-based control parameters. The network classes of OpenIGTLinkMUSiiC improve upon the performance of existing network classes by implementing a multithread architecture. Finally, we present two ultrasound research applications using OpenIGTLink-MUSiiC.


OpenITGLinkMUSiiC: A Standard Communications Protocol for Advanced Ultrasound Research

This document presents an implementation of an algorithm to perform a morphological opening on a graph. The intent is to remove short branches in a graph while preserving the large scale structure. This implementation is based on the algorithm described in ``Efficient Closed Contour Extraction from Range Image's Edge Points''. We have used the data structures from Boost Graph Library (BGL).

The code is available here:
https://github.com/daviddoria/GraphOpening

Morphological Opening on a Graph


This document presents an implementation of an algorithm to find a low edge-count approximation of a complex, discrete, 2D closed contour. This implementation is based on the algorithm described in ``Using Aerial Lidar Data to Segment And Model Buildings'' and ``A Bayesian Approach to Building Footprint Extraction from Aerial LIDAR Data.''

The code is available here:
https://github.com/daviddoria/ClosedLoopSimplification

Closed Loop Simplification

This document presents an implementation of two algorithms, Voronoi Neighbors and Binary Space Partition (BSP) Neighbors. These algorithms find neighbors of a point in a point set that are somehow ``better'' than a ``K nearest neighbors'' or a ``all neighbors within a radius'' query. This type of nearest neighbor query is more computationally expensive, but results in set of neighbors with more desirable properties. The BSP Neighbors search ensures that there is less local duplication, while the Voronoi Neighbors search ensures that the spatial arrangement of the neighbors is as uniform as possible.

These algorithms are explained in ``Point Primitives for Interactive Modeling and Processing of 3D Geometry''.

The code is available here:
https://github.com/daviddoria/SmartNearestNeighbors

Smart Nearest Neighbors

This document describes a contribution to the Insight Toolkit intended to support the process of subdivision of triangle mesh. Four approaches, linear, Loop, modified butterfly and $sqrt{3}$ subdivision schemes were introduced.

Triangle Mesh Subdivision

This document presents an implementation of Moore Neighbor Tracing - an algorithm to find an ordered outline of a blob or contour in an image.

An excellent tutorial on Moore Neighbor Tracing is provided here:
http://www.imageprocessingplace.com/downloads_V3/root_downloads/tutorials/contour_tracing_Abeer_George_Ghuneim/moore.html.


The code is available here:
https://github.com/daviddoria/MooreTracing

Moore Neighbor Tracing

This document presents a VTK implementation of the algorithm described in 
"A clustering method for efficient segmentation of 3D laser data" by Klasing, Klaas Wollherr, Dirk, and Buss, Martin. The algorithm .

The code is available here:
https://github.com/daviddoria/ClusteringSegmentation

Clustering Segmentation for VTK

This document describes a new simple gradient descent optimizer which is a potential
replacement or supplement for RegularStepGradientDescentOptimizer. The optimizer requires only the gradient of the parameters and uses a simple linear model internally. However, it follows the theoretical basis of a trust-region algorithm and is able to achieve greater efficiency on certain image registration cases.


A Gradient Descent Trust Region Optimizer

This code provides an implementation of a simple technique to fill small holes in an image. We use a multi-pass method that fills pixels on the border of a hole with the average of its non-hole 8-connected neighbors. This process is repeated until all holes are filled.

The code is available here: https://github.com/daviddoria/SmallHoleFiller