Transcript (e.g. glaucoma diagnosis and prediction).

Neuro-Fuzzy Glaucoma Diagnosis and Prediction
System
Investigator
Dr. Mihaela Ulieru, Faculty of Engineering, The University of Calgary
CoInvestigator
Dr. Andrew Crichton, Faculty of Medicine, The University of Calgary
Research
team
Dr. Nicolae Varachiu, Cynthia Karanicolas, Mihail Nistor, Faculty of
Engineering, The University of Calgary
Presented papers based in this project
Title
Authors
Title
Authors
IASTED International Conference, Banff, July 2002
Integrated Soft Computing Methodology for Diagnosis and
Prediction with Application to Glaucoma Risk Evaluation.
Mihaela Ulieru, Faculty of Engineering, The University of Calgary
Gerhardt Pogrzeba, President and CEO, TRANSFERTECH GmbH,
Braunschweig, Germany
First IEEE International Conference in Cognitive InfromtaticsICCI’02,
Calgary, August 2002. Computational Intelligence for Medical
Knowledge Acquisition with Application to Glaucoma.
Nicolae Varachiu, Cynthia Karanicolas, Mihaela Ulieru, Faculty of
Engineering, The University of Calgary
Introduction
Diagnosis: to determine if a patient suffers of a specific disease; if so, to
provide a specific treatment
Glaucoma: a progressive eye disease that if left untreated, can lead to
blindness
The main challenge for glaucoma specialists is the evaluation of the risk
for its occurrence and the prediction of disease progression to establish
a suitable follow up and treatment accordingly
Most cases in glaucoma diagnosis are quite evident, but at least 5% of
them will be ambiguous
For these special cases the assessment of an “expert machine” can
be essential in determining the right time for a follow up check as well
as in-between treatment
In response to this need we have developed an integrated diagnosis
and prediction methodology that uses several soft computing
techniques
5
Glaucoma
Cupping of the Optic nerve head
Visual field Loss
Elevated Intraocular Pressure
6
Loss of visual field
Clear image of a road.
Note runner with white shirt on the left.
Glaucoma Visual Field Loss
LEFT EYE
Arc shaped loss of sensitivity starting
from the normal blind spot
(near where the runner is)
into the inside (nasal) field of vision
Glaucoma - severe visual
field loss. Only a small central
island
of vision remains. The centre of
the vision is cut through horizontally
as well
7
Intraocular Pressure
The inner eye pressure (also called
intraocular pressure or IOP) rises
because the correct amount of fluid
can’t drain out of the eye
8
Optic disc nerve damage
9
Glaucoma can also occur as a result of:
An eye injury
Inflammation
Tumor
Advanced cases of cataract
Advanced cases of diabetes
Also by certain drugs (such as steroids)
10
Treatments
Medications
Laser surgery
Filtering
surgery
11
Knowledge representation
Knowledge repository
Fuzzifier
Inputs
Fuzzy logic
Inference
System
(Processing
model)
Defuzzifier
Outputs
12
<x, T(x), U, G, M>
Linguistic variables
x = the Intraocular Pressure (IOP)
T(IOP) = {Low, Normal, High}
U = [0, 45] (measured in mm of Hg)
Low might be interpreted as “a pressure above 0
mm Hg and around 11mm Hg”; Normal as “a
pressure around 16.5 mm Hg” and High as “a
pressure around 21 mm Hg and bellow 45 mm
Hg”.
13
Membership
Function
1
Low
Normal
High
0
0
12
16.5
22
45 mm Hg
Fuzzy sets (linguistic terms: Low, Normal, High) to characterize the linguistic variable
Intraocular Pressure - IOP
14
Knowledge
Acquisition
Iterative process that involves domain expert(s),
knowledge engineers and the computer
15
Knowledge
acquisition steps developing an understanding of the application
domain
determination of knowledge representation
selection, preparation and transformation of data
and prior knowledge
knowledge extraction (machine learning)
model evaluation and refinement
Design of the knowledge engine for disease assessment
The diagnosis of Glaucoma comprises the analysis of a myriad of risk
factors, each of them related to the diagnosis with different degrees.
The rule base is being developed following an incremental development
process
Existing data,
Requirements,goals
Visits to dr.’s office
Ophthalmologist
feedback
Visits to dr.’s office
Ophthalmologist’s
feedback
Neuro – fuzzy
System
Complete set of
fuzzy rules
Top-level
specifications
Incremental
development plan
Iteration 1:
First set of rules
Iteration 2:
Second set of
rules
Iteration n
17
Main steps of the process
Gather and select relevant information to
create or modify the set of rules
Create, add or modify linguistic variables and/or
fuzzy rules
Ophthalmologist’s feedback
Rule set evaluation and refinement
In the first increment a minimal group of Fuzzy IF-THEN rules has been
created. This ‘basic’ set of rules is the foundation for selecting relevant
learning data for improving the prediction engine.
Different risk factors and data is being used to add new rules in each
successive increment.
Each increment will contain all previously developed rules plus some
new ones determined to be relevant by the medical expert.
19
Fuzzy linguistic variables
N°
x
T (x)
U
Low
damage
1
Visual
field
tests
Damage
[0, 76]
Severe
damage
M
A1LD = {0/1 15/1
30/0 76/0}
A1D = {0/0 15/0
30/1 45/1 60/0
76/0}
A1SD = {0/0 45/0
60/1 76/1}
A2N = {20/15/1
20/20/1 20/50/0
20/400/0}
A2A = {20/15/0
20/20/0 20/50/1
20/400/1}
Measure
ment
unit
Low
points
2
Visual
acuity
Normal
Abnormal
[20/15
20/400]
3
Myopia
High
[-10, 4]
A3 = {-10/1 -4/1
0/0 4/0}
No.
4
Cup to
disc
High ratio
[0 1]
A4 = {0/0 1/1}
Number
Number
20
Fuzzy linguistic variables
N°
x
T (x)
U
M
Measure
ment
unit
[0, 45]
A5H = {0/0 16.5/0
22/1 45/1}
A5M = {11/0 16.5/1
22/0}
A5L = {0/1 11/1
16.5/0 45/0}
MmHg
5
IOP
High
Normal
Low
6
Diurnal
Fluctuati
ons of
IOP
Low
High
[0, 10]
A6L = {0/1 5/0 10/0}
A6H = {0/0 3/1
10/1}
MmHg
7
Age
Old
[0, 100]
A7 = {0/0 40/0 80/1
100/1}
Years
old
Output
OL = {0/1 33/1 50/0
100/0}
OM = {0/0 33/0
50/1 66/0 100/0}
OH = {0/0 50/0
66/1 100/1}
8
Risk
Low
Moderate
High
21
Output interpretation
Low risk: follow-up within 6-12 months
Moderate risk: follow-up within next 2-6 months
High risk: follow-up within next few weeks
22
If- Then Rules
23
Example
Visual field tests
Visual acuity
Myopia
Cup to disc
IOP
Diurnal Fluctuations of IOP
Age
FCM Result
Doctor’s action
45
20/150
-9.75
0.8
15
0
80
51.765: next 3-4 months
Appt within 3-4 months
The diagnostic methodology at a glance
Ulieru and Pogrzeba
The methodology has been designed around the software suite
developed by Transfertech GmbH Germany, by integrating several of
their packages.
Aim: emulate the assessment done by the expert physician and collect
relevant data for predicting the disease progression
Diagnosis Engine: embeds expert knowledge
Prediction Engine: developed in a three-step process
Diagnosis
Machine
Parameters
(Measured)
Diagnosis
Engine
Disease
Assessment
Prediction
Engine
Prediction
Prediction
Treatment
Doctor’s
Decision
Follow-up Time
Doctor’s
Decision
Data Base
Machine Parameters
Disease
Assessment
Treatment
Time
Prediction
An evolutionary learning strategy for tuning the
prediction engine
This step assumes a database with sufficient patient information is
already available
The design of the database was a challenging process
Input handwritten patient files.
Database contains: measured parameters, disease assessment,
treatment and time interval decided by medical expert and the result of
the prediction engine.
1. Only once creation of CAM project
Data File (once only exported from Database)
Previous
Disease
Assessment
...
Machine
Parameters
...
Exported
File
Treatment
Time
New Disease
Assessment
...
...
...
Treatment
Time
New Disease
Assessment
Mark
...
...
...
...
Treatment
Time
New Disease
Assessment
Mark
...
...
...
...
New Rule Base
CAM
Fuzzy
Project
Create
CAM
Extention
For
Marking
2. Set Marking using FCM
Database
Machine
Parameters
...
Previous
Disease
Assessment
...
CAM
Fuzzy
Project
Data read by DDE
with
Extention
for
Marking
FCM
3. Learning Stage
Database
Machine
Parameters
...
Previous
Disease
Assessment
...
Old
Fuzzy
Engine
File reading
File
reading
EVO
Uptated Fuzzy Engine
Web-centric extension of the system
Enable data from several clinics to contribute to the knowledge
refinement process.
The prediction system and the central database will be placed on a
central server
Database will be updated periodically
A copy of the diagnosis and prediction engines will function in each clinic
and will be updated after the learning process is done on the central
‘master’ copy
Secure and reliable connection between local engines to the ‘master’
engine
Currently, we are working in the development of a holachy, that would
enable the access of the diagnosis and prediction system from clinics and
by nomadic patients
Conclusions
Our goal is to make this system available on the international health
care arena, therefore several standards have to be investigated and
reconciled (e-health).
The computational intelligence methods increase the accuracy and
consistency of diagnosing, risk evaluation and prognostic of glaucoma
Computational intelligence can embed in a natural way the
uncertainty surrounding the complex medical processes, and in our
specific situation can increase the accuracy and consistency of
diagnosing, risk evaluation and prognostic of glaucoma