Automated Diagnosis of Retinal Images Using Evidential Reasoning

1
Automated Diagnosis of Retinal Images Using
Evidential Reasoning

• Reetal Pai
• March 4, 2002

M. S. Thesis Defense Department of Electrical
Computer Engineering Clemson University
2
Goals
Diagnosis
Retinal Image
Annotation
Eye

• STARE (Structured Analysis of the Retina)
• Measure key features
• Annotate image contents
• Compare manifestations in different images

3
Retinal Image
Optic nerve
Lesions
Tortuous blood vessels
4
Process of Diagnosis
Automated diagnosis of retinal images
Clinical diagnosis of retinal images
Retinal Image
Experts beliefs P(MS/D)
Eye
Automated System
Annotation
System diagnosis
Ground truth diagnosis
5
Influence diagram for our domain
ASR
Coats
PDR
BRVO
Hemi CRVO
Emboli
Diagnoses
CRAO
CNV
BDR
BRAO
Macroaneurism
CRVO
HTR
Artery sheath
Inner retinal infract
Artery oxygen
Artery color
ERM
Cotton wool
Subretinal fibrosis
NVE
CME
Photocoagular scar
A-V change
Vein narrow
VH
On Collateral
Macroaneurism
Artery dilation
Cherry red
TXRD schisis
RPED
Manifestations
Bv Specular .Reflex
On Swelling
Vein color
On Color
Blot Hemorrhage
CNV
Ghost Bv
Emboli
Artery narrow
Macula data
On Hemorrhage
NVD
Preretinal Hemorrhage
Subretinal Hemorrhage
Microaneurism
Vein dilation
Drusen
On Palor
Teleanglecta
Retinal exudates
6
Belief Table

• Experts knowledge summarized
• Probability values

7
Overview of Thesis
Beliefs
Reasoning
System Response
Retinal Image
Evidence Selection
Eye
8
Evidence Selection - Example
ASR
Emboli
Diagnoses
Bv specular reflex
Inner retinal infract
Artery color
Subretinal hemorrhage
Cotton wool
A-V change
Cherry red
Emboli
Blot hemorrhage
Artery oxygen
NVE
VH
Vein narrow
ERM
On Collateral
Retinal exudate
TXRD schisis
Photocoagulator scar
Subretinal fibrosis
On Palor
Preretinal hemorrhage
Teleanglectansis
Manifestations
Macula data
CME
CNV
RPED
On Swelling
Microaneurism
On Color
Macroaneurism
Artery dilation
On hemorrhage
Vein dilation
Artery narrow
Ghost Bv
Artery sheath
Vein color
NVD
Drusen
9
Evidence Selection

• Narrower focus may improve diagnostic ability
• Manifestation may be absent or present
• Specific states for manifestations
  Evidence
• Evidence is individual to each image
• All evidence
• Linked evidence
• Annotated evidence

10
Formulations - Bayes Rule
Bayes Rule
P(B/A) P(A/B)P(B)
P(A)
Applied to our system
J
P(Di/Mj Sk) ? ?j1 P(Mj Sk/Di)?P(Di)
J
I
?j1 ?x1 P(Mj Sk/Dx)
i1..I, j1..J, k1Kj

• Problems
• Presence of zeros in the belief table
• Presence of multiple diseases

11
Formulations - Noisy MAX
k
C
P(M Sk/D) ?c1?n0P(M Sn/Dc) - A
Noisy MAX formulation
A P(M Sk-1/D)
867
P(Di/MSk) ?x1 P(Dx/MSk) if Di ?Dx 0
if Di ?Dx
Recombination

• Problems
• Total number of diagnoses combinations
  867
• Computationally expensive
• Better recombination scheme

12
Relativity of Beliefs
13
Formulations - Normalized sums
P(Mj S/Di) P(Mj S/Di) - mini (P(Mj
S/Di)
Normalization
maxi (P(Mj S/Di) - mini (P(Mj S/Di)
J
P(Di/Mj S) ? ?j1 P(Mj S/Di)
Normalized sums

• Features
• Discards global relativity across
  manifestations
• Maintains local relativity for each
  manifestation
• Novel formulation

14
System diagnoses
Evidence
P(D1/Minput) P(D2/Minput) .. .. .. .. .. .. P(D13/
Minput)
Formulation
System diagnoses
Other diagnoses (not present)
15
Fishers Linear Discriminant
Set A system diagnoses
Set B non system diagnoses
1
0
P(D/M)

• Fishers linear discriminant is calculated for
  each partitioning as follows

• Maximum Fishers gives best separation

16
Evaluation

• Three reasoning formulations are applied in all
  the three contexts of evidence - all, linked and
  annotated
• System performance evaluated in terms of
  diagnostic accuracy
• Two measures of accuracy
• Match
• Perfect match
• Match if GT ? MS ? ?
• Perfect match if GT MS

17
Results

• Results using the match criterion

• Normalized sums outperformed the noisy MAX
• Normalized sums annotated evidence better than
  all evidence
• Noisy max did not show a similar performance
  ratio

18
Results (continued)

• Results using the perfect match criterion

• System is far from optimal
• 23 best system performance - perfect match
• 1 out of 4 times recognizes all

19
Overview of Thesis
Beliefs
Reasoning
System Response
Retinal Image
Evidence Selection
Eye
20
System Responses
Database (354 images)
Known/ familiar images (198)
Unknown/ unfamiliar images (56)
Normal images
Partially familiar images
I dont know
21
Unfamiliar diagnoses

• The threshold system output value 1.2
• System output lt 1.2 , then Unfamiliar image
• System output gt 1.2, then Familiar image

22
Results

• Considerable overlap between the two
  distributions
• System performance on familiar diagnoses has
  fallen from 75 to 60
• All unfamiliar diagnoses are lumped together
• High system performance on partially familiar
  images

23
Normal Case

• Recognize retinal images exhibiting no disease
• 25 of the 38 normal images no abnormalities
• No patterns to the annotations or system output
  values
• Problems in data collection annotation/ground
  truth diagnoses
• Need more images

24
Overview of Thesis
Beliefs
Reasoning
System Response
Retinal Image
Evidence Selection
Eye
25
Frequency versus Belief
Automated diagnosis of retinal images
Retinal Image
Frequencies P(MS/D)
Annotation
Automated System
Ground truth diagnosis
System diagnosis
26
Variation between frequencies and beliefs

• Difference between frequency and belief is less
  than 0.2 in most cases
• The few cases in which this is not true are to be
  studied in future research

27
Results

• Results of these two tests seem contradictory
• Sample size is very small
• Frequency computation is more time and effort
  consuming (took months) than belief acquisition
  (took days)

28
Overview of Thesis
Beliefs
Reasoning
System Response
Retinal Image
Evidence Selection
Eye
29
Case Difficulty
Hypothesis
Database
Familiar images
Medium images
Difficult images
Easy images
30
Case Difficulty
Annotation
Image
Small Easy
Distance
Medium Moderate
Ideal/Separable annotation
Ground truth diagnosis
Large Difficult
31
Image Distances
ck key manifestations absent from image
ideal manifestations
32
Image Class Performance

• Only two classes easy, difficult
• Images with ck lt 0.95 are easy, with ck gt 0.95
  are difficult
• Correct diagnosis requires at least one key
  manifestation

33
Conclusions

• Normalized sums outperforms existing Bayesian
  techniques in our system
• Narrower selection of evidence improves
  diagnostic ability in normalized sums
• Ability to recognize unfamiliar diagnoses leads
  to degradation in system performance
• The discussion on the use of frequencies instead
  of beliefs can be pursued further
• Case difficulty of images contributes to
  diagnostic accuracy