Report of research activities in fuzzy AI and medicine at

Document technical information

Format pdf
Size 116.5 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

DP
RO
OF
Artificial Intelligence in Medicine 578 (2000) 1±7
Report of research activities in fuzzy AI
and medicine at USF CSE
Horia-Nicolai L. Teodorescu*, Abraham Kandel, Lawrence O. Hall
University of South Florida, Computer Science and Engineering (CSEE),
4202 E. Fowler Ave., Tampa, Fl 33620, USA
TE
Received 28 February 2000; received in revised form 12 July 2000; accepted 01 August 2000
Abstract
EC
Several projects involving the use of fuzzy and neuro-fuzzy methods in medical applications,
developed by members of the Department of Computer Science and Engineering, University of
South Florida, Tampa, Florida, are brie¯y reviewed. The successful applications are emphasized.
# 2000 Elsevier Science B.V. All rights reserved.
Keywords: Neuro-fuzzy system; Sudden infant death syndrome; Fuzzy logic
1. Introduction
2. Mission statement
CO
RR
The Department of Computer Science and Engineering (CSEE) at University of South
Florida includes several scientists working in the ®eld of fuzzy and neuro-fuzzy tools such
as fuzzy clustering, fuzzy ®ltering, fuzzy image segmentation, fuzzy expert systems with
medical applications, including diagnosis, monitoring, and rehabilitation. Recently, a
research group, named Neuro-Fuzzy Systems in Human-Related Sciences Group was
created. The group founding members are Kevin Browyer, Dmitry Goldgof, Larry Hall,
Abraham Kandel, and Horia-Nicolai Teodorescu (currently the group manager). The
Group organization is ¯exible, with a change of management taking place every 6 months.
Our research group was established to foster research, teaching, and academic cooperation in the ®eld of neuro-fuzzy systems in human-related ®elds, namely in domains like
*
UN
Corresponding author. Fax: ‡1-813-974-5456
E-mail address: [email protected] (H.-N.L. Teodorescu).
0933-3657/00/$ ± see front matter # 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 3 3 - 3 6 5 7 ( 0 0 ) 0 0 0 8 3 - X
DP
RO
OF
2
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
medical sciences, biology, vision, speech processing, linguistics, psychology, and cognitive sciences. Based on the expertise of its individuals, the group is dedicated to the
development of new fundamental knowledge, processes or procedures through research,
promoting multi-disciplinary approaches. It is a claimed objective of the group to fuse
various approaches, including fuzzy logic and fuzzy algebraic tools, hybrid neural networks, pattern recognition, non-linear dynamics, symbolic systems, and machine learning.
Moreover, the group aims to integrating AI into various medical ®elds, including diagnosis,
prostheses, image processing, rehabilitation and monitoring of patients. Also, the group
aims to foster education in these ®elds.
3. Projects
EC
TE
Some of the recently carried projects include applications in medical fuzzy expert
systems, image processing (mainly segmentation, in the frame of an AI in human brain
segmentation project, and contour interpolation/data compression), respiration monitoring
and respiratory event warning, tremor assessment, modeling of biological processes, and
speech processing in view of medical applications. The projects were carried as, internal
projects, independent research of some of the members (supported by start-up companies),
or as international co-operation, mainly with the Romanian Academy and the Swiss
Federal Institute of Technology, Lausanne, Switzerland. In this paper we do not report on
all current or past researches performed in this department and the selection we performed
is largely based on data availability.
3.1. Use of fuzzy methods in segmentation of tumor images [1±3]
UN
CO
RR
Several methods, based on fuzzy information and knowledge in image segmentation,
have been developed to detect brain and breast tumors. We brie¯y present only one of these
methods. Further information can be obtained from the quoted resources and references.
A knowledge-based system integrated with unsupervised fuzzy clustering to automatically segment and label tumors in magnetic resonance slices of the human brain was
developed. Slices are initially segmented by an unsupervised fuzzy c-means algorithm. The
rule-based system uses model-based recognition techniques and further fuzzy clustering to
iteratively locate tissues of interest. The system uses knowledge obtained during preprocessing. Further fuzzy re-clustering is aided by the use of initialization and training data
created by the knowledge-based system. Over-clustering plays an important role. Providing more clusters than there are tissue types, the amount of under-segmentation was
reduced and objects of interest are easier to ®nd. The advantages of the proposed approach
are demonstrated by the successful performance of the system, and the ®nal segmentation
of Glioblastoma multiforme tumor compared favorably with hand-labeled images of the
tumor.
In another approach, using fuzzy rules for segmentation, fuzzy rules for partially
segmenting MR images of the brain were built to operate on density weighted intensity
feature images. The tissue thresholds determine the antecedent fuzzy sets of the rules. The
thresholds were found via histogram analysis applied to each image slice to which the rules
DP
RO
OF
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
3
TE
will be applied. The turning points in the histograms are essentially the approximate
boundaries between tissue types. The turning points (peaks, valleys or the beginning of a
hill) are automatically chosen on each test slice. From the turning points in the histogram,
fuzzy rules to identify tissue classes such as white matter, gray matter, or bone, are
generated. The rules and antecedent fuzzy sets were generated by examining the intersection of tissue types in the three intensity histograms. Based on the rules, the voxels are
classi®ed. An unclassi®ed voxel (i.e. having a zero membership in all classes) is assigned a
membership that is the average membership of its neighbors. If a voxel has a membership
of 1.0 in a class A, while surrounding voxels have zero membership in that class, then the
isolated voxel's membership to A is zeroed. This step is aimed at reducing classi®cation
errors. Finally, the voxel memberships in all classes are normalized to 1. The voxels that
belong to classes with memberships greater than 0.8 are generally correctly assigned. The
rest of the voxels are more problematic. Hence, we re-group them with a semi-supervised
clustering algorithm, ssFCM. The voxels with membership greater than 0.8 are used as
training voxels for ssFCM. The ssFCM algorithm works as fuzzy c-means (FCM) except
that training voxels cannot change clusters and will always in¯uence the cluster centroid to
which they are assigned. The method has demonstrated advantages over other methods.
More details can be found at (http://morden.csee.usf.edu/hall/adrules/segment.html).
3.2. Algebraic neuro-fuzzy systems for image processing [4,5]
UN
CO
RR
EC
In this project aimed to develop and test a new tool, namely algebraic fuzzy neural
networks, in applications like signal ®ltering, classi®cation, data interpolation and
information compression. The algebraic fuzzy neural networks represent neural network
with fuzzy neurons, the operations performed by the neurons being fuzzy algebraic
operations, instead of the classic fuzzy logic operations. The weights (synapses) are
represented by fuzzy numbers. Both sigmoidal and RBF (radial-basis function) neurons
have been used. The input fuzzy numbers are either triangular or trapezoidal. The main
dif®culty in applying algebraic fuzzy neural networks is related to the training algorithm.
Several algorithms that work under algebraic fuzzy operations have been developed and
successfully demonstrated. The medical applications addressed included non-linear signal
processing, and image processing. An example is the approximation of the blood microvessel internal diameter and wall thickness, based on microscopy images. The application
consists in blood vessel image reconstruction and, subsequently, the prediction of ¯ow
variations along the vessel. The use of fuzzy methods is required by the imprecision related
to the vessel boundary in the microscopic images. The technique has potential applications
in the diagnosis of the circulatory system and in blood vessel re-constructive surgery.
Solving the same task manually is tedious and expansive. Images were obtained based on
2D slices made at equal distances. The image is processed by conversion to gray levels and
then the contrast is enhanced by classic techniques. After the selection of the regions
including the blood vessel of interest, the image is processed in view to extract the
information on the wall thickness and lumen (internal diameter). The wall is approximated
in every point by a triangular fuzzy set representing the thickness. A sample image with the
related explanation on de®ning the fuzzy thickness and diameters is shown in Fig. 1. These
theoretical and software developments and applications were reported in several papers
DP
RO
OF
4
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
TE
Fig. 1. Image of a section in a micro-vessel and the fuzzi®cation of the thickness of the blood vessel wall. The
uncertainty about the boundaries is accounted for by the use of a triangular fuzzy number (after an internal
report, seminar for fuzzy systems ``Gr. C. Moisil'').
EC
and partly summarized in a chapter [4]. This research was performed in co-operation with
Dr. D. Arotaritei (Institute of Computer Science of the Romanian Academy) and the partial
support of the Romanian Academy
3.3. Fuzzy fusing of classic and chaotic numerical parameters in a hand tremor
rehabilitation system [6]
UN
CO
RR
We developed a comprehensive methodology for tremor analysis, including classic and
chaotic parameters, where the non-linear analysis complements the time±frequency
analysis. We evidenced that tremor includes a signi®cant chaotic component, and this
component may be of interest in diagnosis and rehabilitation. Various practical aspects,
from sensors to feedback procedures based on fuzzy logic, were solved and results
incorporated in equipment and related software. Moreover, an experimental arrangement
and dedicated software for tremor analysis and feedback for rehabilitation purposes was
developed. Non-linear data analysis and a fuzzy method to process the signal, to assess the
type of tremor, moreover to provide an easy to grasp feedback for rehabilitation purpose,
were incorporated. The system for rehabilitation is based on fuzzy assessment of fused
features of the tremor movements. Both classic and non-linear indices were fused into a
few, easy to represent and grasp fuzzy measures. A neuro-fuzzy system for tremor control
was also devised. The feedback should be easy to understand and learn, and should make
use visual and audible information. To make it easy to learn, the information has to be
appropriately compressed.
With the purpose of creating a feedback to help patients to become aware of the
characteristics of the tremor of their limbs, moreover to help them controlling the tremor,
we used a representation of the tremor by images and sounds. This representation is
synthetic and is based on the linguistic and fuzzy valuation of the main parameters of the
DP
RO
OF
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
5
tremor. The use of linguistic and fuzzy valuations is justi®ed by its simplicity (easy to
understand by the patient), moreover by good information compression. The rules used to
globally characterize the tremor take into account, among others, the tremor amplitude,
main frequency, ratio low-to-high frequency power content, the correlation dimension and
the irregularity of the signal. The rules establish the relations between the pattern in the
space of the tremor signal space and the classes of the tremor signal. The results of the
inference are defuzzi®ed and used in the feedback. There are various feedback facilities
under current evaluation with the system. Most of the research related to this topic has been
supported by Techniques & Technologies Ltd. and by the Grant 7RUPJ-48689 from Swiss
Research Founds (FNS), Switzerland.
3.4. Applications to monitoring the respiration of infants: SIDS prevention [7]
3.5. Other projects
CO
RR
EC
TE
This application was developed in the frame of a start-up (Sensitive Technologies LLC).
The system, based on two patents [7] is aimed to monitor infants aged 0 to 24 months that
may be at risk of sudden infant death syndrome (SIDS), yet are living a normal life within
their family. The dif®culty of the problem relates to the restrictions on sensors (they should
be completely non-obtrusive and should collect data without contact) and to the need to
make a highly reliable alarming. Previously known systems for monitoring either
respiratory signals, or movements of a subject during sleep and especially for detecting
SIDS-related apnea typically employ accelerometers and simple electronics or sleep
analysis software for analyzing data received from sensors. Such systems have shortcomings, including unreliability in event detection, largely due to inability to distinguish
between respiratory movements and other movements.
Beyond using a novel type of sensor, the system developed for this product uses several
sensors, and a data fusing, pattern recognition system based on neuro-fuzzy techniques.
The monitoring system includes a sensor for detecting the respiration and a second sensor
for detecting the presence and movement of the infant or proximal objects around the crib.
A third acceleration sensor helps to discriminate between the movements of the crib frame
induced by external factors and movements induced by respiration. Signal from the sensors
are processed to extract respiration- and non-respiration-related signals and the respirationrelated signal patterns are compared to normal patterns. The system fuses the data from
various sensors and makes aggregated decisions to generate speci®c warnings for speci®c
respiratory events. The overall processing is able to take into account ``secondary'' factors
empirically known to in¯uence the respiration stop risk. Such factors are the ambient
temperature, SIDS-related family history of the infant, age of the infant, sex, birth-weight,
drug therapy history, infectious state of the infant, mother's age and smoking status of the
mother. To cope with the large amount of data, the uncertainty about the data, and the wide
range of circumstances that in¯uence the decision, the decision process is based on a fuzzy
knowledge-base. Various levels of warnings are available.
UN
An interface system has been developed for creating a feedback by vibro-tactile stimuli
and images to help correct pronunciation, with the aim to offer a feedback during training
DP
RO
OF
6
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
of hearing disabled children. The assessment of the pronunciation involved a fuzzy
knowledge-base. The device converts speech-related information into visual and tactile
information.
In another project, models for biological processes have been developed based on
chaotic neuro-fuzzy systems [8].
4. Future trends
TE
We believe that a better understanding of the non-linear properties of fuzzy and neurofuzzy systems with defuzzi®cation, and of the dynamics of fuzzy iterated processes is
essential for the evolution of the ®eld. At the same time, the improved understanding and
theoretical foundations of non-linear methods in medicine condition the use of fuzzy and
neuro-fuzzy systems in medicine1. Both ®elds lack yet the desirable theoretical development.
Our current and planned projects include the following.
CO
References
RR
EC
Further application of neuro-fuzzy methods to medical image processing.
Further application of neuro-fuzzy systems in control of prostheses and use of artificial
life (AL) technologies, including evolvable hardware for neuro-fuzzy systems.
A comparative analysis of dynamical neuro-fuzzy systems in modeling applications is
under progress. The analysis is aimed to throw light on some new concepts related to
chaos in neuro-fuzzy systems, and similarities to biologic, economic and social
processes.
An analysis of errors in hardware implementations of neuro-fuzzy systems is under
progress. The aim is to determine which solutions are better suited for specific
applications.
A tentative project refers to the analysis of the dynamics associated to a knowledge
processing systems. The concept of dynamics associated to a knowledge processing
systems (KPSs) is addressed in the context of medical applications. The dynamic
analysis is specifically applied to fuzzy KPSs, including fuzzy knowledge bases
(FKBs), fuzzy decision-making and decision support systems (FDSs) and fuzzy expert
systems (FESs). Applications envisaged range from diagnosis to treatment to anesthesia
control to nutrition. A testing methodology aiming to assess the expected dynamic
behavior is presented.
[1] Clark M, Hall L, Goldgof D, Silberger M. In: Cabonell, JG, Siekman J, editors. Using Fuzzy Information in
Knowledge Guided Segmentation of Brain Tumors, in Fuzzy Logic in AI: Towards Intelligent Systems,
Lecture notes in AI 1188. Springer, 1997. p. 167±81.
1
UN
The volumes given in ``Further reading'' section in bibliography may interest the reader looking for
applications of fuzzy and neuro-fuzzy systems in medicine. These volumes include chapters on several
significant projects around the world involving fuzzy and neuro-fuzzy systems in medical applications. The web
pages related to tour department including links to several papers are also given in this section.
DP
RO
OF
H.-N.L. Teodorescu et al. / Artificial Intelligence in Medicine 578 (2000) 1±7
7
TE
[2] Clark MC, Hall LO, Goldgof DB, Clarke LP, Velthuizen R, Silberger M. MRI Segmentation using Fuzzy
Clustering Techniques: Integrating Knowledge IEEE Engineering in Medicine and Biology Magazine,
Special Issue on Fuzzy Logic in Medicine. vol. 13, no. 5, November/December 1994, p. 730±42.
[3] Clark MC, Hall LO, Goldgof DB, Velthuizen R, Murtagh R, Silbiger MS. In: Teodorescu HN, Kandel A,
Jain LC, editors. Unsupervised Brain Tumor Segmentation using Knowledge-Based Fuzzy Techniques,
Fuzzy and Neuro-Fuzzy Systems in Medicine. 1998. p. 137±69.
[4] Teodorescu HN, Arotaritei D. Algebraic Neuro-Fuzzy Systems and Applications. p. 193±238 [chapter 7].
[5] Teodorescu, HN, Arotaritei D. In: Proceedings of The Seventh International Fuzzy Systems Association
World Congress IFSA'97 on Analysis of Learning Algorithm for Algebraic Fuzzy Neural Networks,
Prague, Czech Republic, vol. IV,1997 June 25±29, p. 468±73.
[6] Teodorescu HN, Kandel A. Non-linear analysis of tremor and applications. Jpn J Biomed Eng 13 (5),
pp. 11±20 May, 1999
[7] Teodorescu H, et al. Respiration and movement monitoring system. U S Patent No. 6,011,477, Jan. 4, 2000.
Original application 60/053,543, July 23, 1997, ®nal application 09/004.108, January 7, 1998. (Search the
http://www.uspto.gov/patft/index.html data base of the US Patent And Trademark Of®ce using patent
number or author). Related patents and patent applications are: HN Teodorescu, Position and movement
resonant sensor. U S Patent No. 5,986,549. Nov. 16, 1999 WO9905476 (World Patent Application), HN
Teodorescu: Position and movement resonant sensor. International/European 1999-02-04 AU8641498
(equivalent to WO9905476) WO9904691 (World Patent), HN Teodorescu, D Mlynek: Respiration and
movement monitoring system. 1999-02-04.
[8] Teodorescu HN, Kandel A, Brezulianu A. Biologic Dynamic Processes Modeling Based on Chaotic Fuzzy
Systems Biomedical Soft Computing and Human Sciences (Japan), vol. 4, no. 1, pp. 1-10, 1998
EC
Further reading
UN
CO
RR
[9] Teodorescu HN, Kandel A, Jain LC, editors. Fuzzy and Neuro-fuzzy Systems in Medicine. CRC Press,
Florida, USA, p. 394, ISBN0-8493-9806-1, 1998. (Also see Highly Recommended Resources list of
NEUROLIFE (http:www.omega23.com/books/s5/neurolife).
[10] Teodorescu HN, Mlynek D, Kandel A, Zimmermann HJ, editors. Intelligent Systems and Interfaces.
Kluwer. ISBN 0-7923-7763-X, February 2000, p. 480.
[11] Teodorescu HN, Kandel A, Jain LC, editors. Soft-Computing in Human-Related Sciences. CRC Press,
Florida, USA May 1999 ISBN 0849316359, p. 370.
[12] Teodorescu HN, Jain LC, editors. Intelligent Technologies for Rehabilitation. CRC Press, Florida (To
appear, summer 2000).
[13] http://morden.csee.usf.edu/ailab/hall.html Web page of the laboratory for AI/Dr. Hall; links to several
papers on fuzzy logic in image processing. http://morden.csee.usf.edu/KB-Papers/IJCAI/ijcai.html is a
page related to the above, presenting some easy to follow explanations on fuzzy techniques in image
processing.
[14] http://marathon.csee.usf.edu/kwb/medical-imaging.html Announcement of a Mammography Image
Analysis Research Database.
[15] http://marathon.csee.usf.edu/ The Computer Vision/Image Analysis Research Laboratory at the University
of South Florida (Dr. Goldgof).
[16] http://www.Sensitivetech.com/ Web page of Sensitive Technologies LLC, USA; presents basics of the
technology used in infants monitoring for SIDS prevention.

Similar documents

×

Report this document