robust cerebrum and cerebellum segmentation for neuroimage analysis

NA-MICNational Alliance for Medical Image Computing http://na-mic.org

National Alliance for Medical Image Computing http://na-mic.org

Robust Cerebrum and Cerebellum Segmentation for Neuroimage Analysis

Jerry L. Prince, Aaron Carass, Lotta Ellingsen,

and Min Chen

Johns Hopkins University

Ron Kikinis and Nicole Aucoin

Brigham and Women’s Hospital

1

National Institute on Biomedical Imaging and Bioengineering 1 R21 EB009900-01


“Skull Stripping”

• Isolate the brain in MR images

• Existing algorithms yield inconsistent results

human rater is still gold standard

• SPECTRE is a “hybrid” technique (top-down and bottom-up) that preserves cortical gray matter and yields high Dice coefficients

Typical SPECTRE result

SPECTRE Result on patient

2


Features

• Do not be overly aggressive:– affects quantification (thickness)

of cortical gray matter– affects determination of pial

layer

• Provide added flexibility:– separate hemispheres,

cerebrum, cerebellum and brain stem

• Rigorously validate:– evaluate on thousands of highly

varied data sets– compare with existing

algorithms

3

Manual Result

Some cortex is “stripped” away


Project Overview

Goal: Develop and augment the SPECTRE algorithm as a module within the NA-MIC Kit software environment—e.g., 3DSlicer. (Funded NIBIB R21)

SPECTRE = Simple Paradigm for Extra Cerebral Tissue REmoval

Specific aims:

1. Rewrite SPECTRE using the Free Open Source Software Development Methodology of NA-MIC. [Hopkins, BWH, Mostly Year 1]*

2. Complete the algorithms necessary for isolating and establishing a coordinate system on the cerebellum. [Hopkins, Years 1 and 2]

3. Parameters will be optimized in order for SPECTRE to perform well with a variety of differently acquired data. [Hopkins, Years 1 and 2]

4. Compare SPECTRE against all available publicly available (source or binary) skullstripping software. [Hopkins, BWH, Mostly Year 2]

*Or: Make SPECTRE a module within 3DSlicer

4


SPECTRE Basics

• A watershed principle is at the core of SPECTRE– Works according to intensity

ordering of WM, GM, CSF in MR images

• Initial masks are found by multiple atlas registration– cerebrum, cerebellum, brain

stem, and whole brain

• classification, morphology, topology, and watershed algorithms are then applied– sanity checks are applied

5


History and Status

• SPECTRE was originally implemented in C/C++– ISBI conference paper in 2007

• SPECTRE was ported to JAVA in order to operate within the Java Image Science Toolkit (JIST environment)– JIST is an open source pipeline and medical image processing

environment: http://www.nitrc.org/projects/jist/

• Further development and testing has been carried out on the Java code within JIST – a journal paper is in the final preparation stage

6


Algorithm Progress

• Cerebellum isolation • Hemisphere separation

7


Performance Evaluation

• Data sets and results– 25 older subjects, cerebrum,

two raters– 3 subjects, repeat scans– 1046 scans, older subjects,

cerebrum, one rater

8

Worst result


What is JIST?

• It is a visual interface, pipeline tool, and collection of 3D image analysis modules

• It works in concert with MIPAV

• Modules can run from a command line interface

• Java platform independent

• Designed for large-scale runs

9


Early Integration Status• Java SPECTRE requires Java

Runtime environment (JRE) and JIST and MIPAV

3DSlicer is being extended to load JRE [BWH]

SPECTRE and libraries are being rolled up in a comprehensive JAR file [JHU]

• The SPECTRE module is described by an XML file that is scanned by 3DSlicer and incorporated as a command line module

XML descriptor is being written [JHU]

10


SPECTRE 2009• Creation of SPECTRE 2009

graphical interface in 3DSlicer is automatic– Required JIST CLI mods

– Done this week, shown in right

• Limitations– no shared memory between

SPECTRE and 3DSlicer

– memory requirements overall are large: 1-1.5GB for SPECTRE

• Benefits– Mechanism permits other modules

to become available to 3DSlicer

– Future SPECTRE development works in JIST, MIPAV, and NA-MIC

11


Future Work

• Release initial version as 3DSlicer plugin– Complete CLI support for JIST/Java in 3DSlicer– Fix Windows Java problem

• Incorporate MGDM for joint cerebrum, cerebellum, and brainstem isolation – Validate MGDM in Java– Augment OASIS atlases with additional labels

• Carry out extensive validations– cerebellum, 20 subject repeat scans, tumor cases

• Optimize SPECTRE speed, flexibility, visualization interactivity

12

robust cerebrum and cerebellum segmentation for neuroimage analysis

Documents