IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 11, NO. 1, JANUARY 2006 47
Integrated Support for Medical Image Analysis
Methods: From Development to Clinical Application
Sı´lvia D. Olabarriaga, Jeroen G. Snel, Charl P. Botha, Member, IEEE, and Robert G. Belleman
Abstract—Computer-aided image analysis is becoming increas-
ingly important to efficiently and safely handle large amounts of
high-resolution images generated by advanced medical imaging
devices. The development of medical image analysis (MIA) soft-
ware with the required properties for clinical application, however,
is difficult and labor-intensive. Such development should be sup-
ported by systems providing scalable computational capacity and
storage space, as well as information management facilities. This
paper describes the properties of distributed systems to support
and facilitate the development, evaluation, and clinical application
of MIA methods. First, the main characteristics of existing systems
are presented. Then, the phases in a method’s lifecycle are an-
alyzed (development, parameter optimization, evaluation, clinical
routine), identifying the types of users, tasks, and related computa-
tional issues. A scenario is described where all tasks are performed
with the aid of computational tools integrated into an ideal sup-
porting environment. The requirements for this environment are
described, proposing a grid-oriented paradigm that emphasizes
virtual collaboration among users, pieces of software, and devices
distributed among geographically dispersed healthcare, research,
and development enterprises. Finally, the characteristics of the ex-
isting systems are analyzed according to these requirements. The
proposed requirements offer a useful framework to evaluate, com-
pare, and improve the existing systems that support MIA develop-
ment.
Index Terms—Distributed computing, grid computing, medical
image analysis, problem-solving environment.
I. INTRODUCTION
M EDICAL images are the basis for a large number of clin-ical tasks in the daily routine of healthcare. Continued
developments in acquisition technology enable capturing the
increasing amounts of high-resolution images that reveal dif-
ferent aspects of the human body’s structure and function with
unprecedented detail. In the clinical routine, such large amounts
of data raise not only storage issues but also challenges for im-
age analysis. In summary, how can/should such large amounts
of data be interpreted efficiently and safely? Adopting compu-
tational aid for medical image analysis (MIA) is no longer an
option, but a necessity.
Manuscript received September 30, 2005; revised February 22, 2006. This
work was supported by a BSIK Grant from the Dutch Ministry of Education,
Culture and Science (OC&W) and is part of the ICT innovation program of the
Ministry of Economic Affairs (EZ). This work was carried out in the context of
the Virtual Laboratory for e-Science project (www.vl-e.nl).
S. D. Olabarriaga and R. G. Belleman are with the Informatics Insti-
tute, University of Amsterdam, 1098 SJ Amsterdam, The Netherlands (e-mail:
silvia@science.uva.nl).
C. Botha is with the Faculty of Electrical Engineering, Mathmatics and Com-
puter Science, Delft University of Technology, 2600 GA Delft, The Netherlands.
J. G. Snel is with the Academic Medical Center, University of Amsterdam,
1100 DD Amsterdam, The Netherlands.
Digital Object Identifier 10.1109/TITB.2006.874929
MIA is an active area of research that aims at the develop-
ment of (highly autonomous) computational methods for image
enhancement, segmentation, registration, measurement, and in-
teractive visualization. Methods are designed to (automatically)
enhance, detect, select, and display features of interest in the
image, with goals such as shortening or eliminating examina-
tion times, reducing subjectivity, and facilitating measurement.
Methods can originate from new algorithms and from innova-
tive combinations of existing ones. In practice, MIA methods
are often implemented as a composition of processing steps that
progressively analyze the image to generate the required result.
The images can belong to sets of different scans and the results
can be of varied data types, such as images, measured values,
or geometrical descriptions. In this paper, we refer to single
or composed methods that perform any type of MIA operation
simply as MIA methods.
Although many MIA methods are proposed in the literature,
only a limited number prove themselves in producing results
with sufficient reliability to be acceptable for application in clin-
ical tasks. This is due to three main reasons. First, it is difficult
to mimic, by software, the performance of trained human oper-
ators in the task of image interpretation. Algorithm design, as
well as finding the best combination of image analysis steps, are
often challenging tasks. Moreover, even when the MIA method
is well established, it might need adaptation when the imag-
ing protocol changes. In many cases, MIA methods simply do
not meet the required quality standards when confronted with
broader settings and are left unused. Second, the validation of a
new method, before it can be used in practice, demands a huge
effort. This is partly due to the high requirements on reliabil-
ity inherent to clinical tasks, and also to the large biological
variability and the lack of ground truth for the objective perfor-
mance evaluation. The quality of the produced results is usually
estimated in a statistical manner, based on a large number of
images and often involving human intervention. Such evalu-
ation studies consume large computational resources for data
storage and processing, which turns out to be a major burden
in the validation of software for MIA. And finally, integrating
a new method into the clinical environment requires its imple-
mentation in workstations that are integrated into the existing
image analysis workflow. This integration might be impossible
to achieve in turnkey systems without vendor participation, or
it could create an uncomfortable situation for the operator, who
needs to move physically among dispersed devices.
Some of the problems discussed above can be reduced when
an adequate (computational) infrastructure is adopted to effi-
ciently handle the logistics of MIA development and deploy-
ment. Problem-solving environments (PSEs) [1], [2], which are
1089-7771/$25.00 © 2007 IEEE