The differential diagnosis of vertical strabismus squint

1
The differential diagnosis of vertical strabismus
(squint) using an Artificially Intelligent
expert StrabNet A.C.Fisher , A.Chandna,
D.Stone, I.Cunningham Dept. of Clinical
Engineering, and St. Pauls Eye Unit, RLUH Dept.
of Paediatric Ophthalmology, Royal Liverpool
Childrens Hospital, (Alder Hey)
(a.c.fisher_at_liv.ac.uk)
Introduction The differential diagnosis of
strabismus is not always straight-forward and
requires the expert interpretation of the
clinical examination by the ophthalmologist or
orthoptist, the clinical expert. The StrabNet
Project develops an artificially intelligent
machine expert (ie. a computer) to classify
vertical strabismus on the basis of learnt rules
solicited indirectly from the clinical
experts. StrabNet is available on the Internet as
StrabNet.com or as a stand-alone program. The
differential diagnosis of vertical strabismus is
routinely performed in clinic using the Prism
Cover Test (PCT, see Figure 1). The magnitude of
the deviation in each of 10 cardinal positions of
gaze is recorded as the power of the neutralising
prism. Prism powers are typically in the range
/- 20 prism dioptres (?D) at 0.5 ?D resolution.
The number of possible combinations and
permutations is enormous. Consequently, the
differential diagnosis into one of 8 types
normally requires the interpretation of these
data by an experienced clinical expert.
build 8 artificial neural networks (ANNs)
Method continued
Clinical experts create 400 artificial PCT
tests 8 groups of 50 for each diagnosis (Classes
1 - 8)
100 random records
learn
50 random records
validate
50 random records
test
1c
1a
1b
Figure 3 Developing the artificially intelligent
expert
Figure 1 a b Prism Cover Test (PCT)
with prism bar and cover (a) un-covered (b) c.
an example of a vertical squint d. nine positions
of gaze the cardinal positions used in the PCT
4a
1d
Graphic Representation of Strabismus
Method A model for representing motor and
sensory aspects of strabismus was designed to
include essential aspects of assessment of
strabismus into a single graphic. The starting
point for the development of the model was the
commonly accepted graphic representation of
strabismus (Figure 4a). An artificial dataset of
400 PCT examinations of 10 measurement fields was
created. These were classified by the Clinical
Experts (AC DS) into 8 classes of strabismus
(Figure 2). An artificial neural network (ANN)
was constructed using the MatLab programming
environment to classify 100 randomly-selected
records (the Training set) into each of the 8
conditions. The remaining data were further
sampled into 2 equal sets of 50 as Validation and
Test sets (Figure 3)
4b
User selects prism values in the range -20 to 20
?D from the Prism Cover Test
4c
When all 10 values are entered, StrabNet responds
with the differential diagnosis
Figure 4 Using StrabNet
Results StrabNet, a simple ANN based on a
3-layer perceptron with back-propagation
(MLP_bp), was 100 accurate in classifying both
the Validation and Test data. This performance
was consistent over 8 consecutive design trials.
Conclusions A relatively unsophisticated expert
ANN model was developed as a machine-based expert
to classify vertical deviation strabismus into 8
diagnostic groups from PCT data. Further
development will include the description of the
minimum dataset required for this model (i.e..
identification by Principal Component Analysis of
redundant data fields (PCT measurements).