Fuzzy Application Library/Technical Applications/Monitoring Glaucoma

Monitoring Glaucoma by Means of a NeuroFuzzy Classifier

by Gudrun Zahlmann and Matthias Scherf of GSF, National Research Center for Environment and Health, medis, Ingolstädter Landstr. 1, D-85764 Neuherberg, e-mail: zahlmann@gsf.de, Phone: +49-89-3187-2347, Fax: +49-89-3187-3370; and Aharon Wegner of Department of Ophthalmology, Clinic rechts der Isar, Technical University of Munich, Ismaninger Str. 22, D-81675 Munich

Citation Reference: The Fuzzy Logic Application Note series is published by Inform Software Corporation on its Internet server to promote the use of fuzzy logic technologies in applications.

Monitoring glaucoma related changes within the eye status of a patient requires a decision whether or not the actual ophthalmic data set represents critical or suspicious situations as a differential diagnosis and a detection of changes within the status over time. This paper describes our approach supporting the differential diagnostic problem. The decision, which type of situation occurs is made by a neuro-fuzzy classifier. The neural net part is based on a RBF network. Fuzzy classification is realised by a fuzzy rule set combining all patient data with the classification results of the neural net classifier to the final decision.


Glaucoma is one of the most severe eye diseases according to the number of blindness cases in the western industrial countries [1]. Therefore an early detection of glaucomatous changes within the patient's eye status is very important. This can be done supporting the decision finding process of primary care settings using approaches of 'intelligent monitors'. The main function of the described glaucoma monitor is to react as a 'watch dog', 'barking' if critical or suspicious situations are detected, otherwise to rely in a 'sleeping' condition. Both cases require the assessment and classification of new data sets. This classification has to be done in two different ways, as a differential diagnosis decision and an evaluation of time-dependent changes. This paper describes the classification methods leading to a differential diagnosis decision.

Input Data

The decision of medical experts about glaucomatous changes of a patient's eye rely on the following data sets:

To the direct measurable data belong the intraocular pressure (IOP) given as a real number in a pressure unit (or as data set of several IOP values as a daily profile) and perimetry data sets. The latter data describe the status of the visual field of the patient measured by special devices (perimeters), which detect the loss in light sensitivity at different stimulus points of the retina (number and locations depend on the applied perimeter type and measuring regimen). These locations are stimulated by flash light of different intensities at different locations within a hemisphere with a background illumination while the patient is fixating to the center. The 'answer' whether or not a stimulus could be seen is given pressing a response button by the patient. Therefore, perimetry is a subjective measuring method in general.

Another input to the decision process are papilla descriptions. These are mainly made by watching the papilla on-line during the eye examination and estimating several parameters like the cup-disc-ratio (CDR), the location of the excavation or a comparison of the CDR's of the right and left eyes. The CDR and the right-left-difference are given as real numbers (estimated), the location is given in linguistic terms like 'central, inferior, superior'. The CDRs and the differences are transformed by the medical experts into a meta level of classification, evaluating them as 'normal, increased' etc.

The third information source are the patient's own reports which are mainly given verbally and unstructured but contain valuable information about the status at the time point of visiting the ophthalmologist and about changes in this status.

Based on these data the ophthalmologist has to decide, whether or not these data and the related findings (transformation to the decision level - situation description at the knowledge level [2]) belong to glaucomatous findings or not and additionally what type of reaction is required. Reactions here can be the decision about a shorter or longer interval to the next examination, a medication or a decision about a type of eye surgery (depending on the glaucoma type, like open angle glaucoma etc.).

Our goal is to model this decision process giving finally a decision support to the ophthalmologist. The final decision maker remains the physician. Data sources and types has led us to fuzzy rules and neural networks as classification methods.

Neural Networks

Due to the fact that it is very difficult to give an exact formal description about how to classify perimetry data we will use artificial neural networks (ANN) to learn the classification task on the basis of preclassified perimetry samples.

Classification tree and ANN classifier approach Figure 1: Classification tree and ANN classifier approach (large)

The design of the ANN classifier is based on the classification tree which is shown in Figure 1 (left side). The motivation in modelling the classification tree is to introduce several decision stages, ranging from rather crude decisions like the 'normal'/'pathological' classification up to refined decisions, like the 'questionable'/'probably glaucomatous' classification. Our approach in modelling the ANN-perimetry-classifier is to design a specialised ANN for every decision level from the root to the leaves of the classification tree respectively (see Figure 1 right).

The specialisation takes place in training every ANN exclusively on a part of the feature space which is defined by the set of perimetry samples according to their classification task. Furthermore the input dimensionality of every ANN is reduced. A perimetry data sample is classified by ANN 1 through ANN 6 respectively. Their output values are interpreted top down according to the classification tree. If for example the ANN 2 classifier gives the result that the perimetry is more 'normal' than 'pathological', we will give more attention to the results of the ANN 3 classifier than to the results of the ANN 4, ANN 5 or ANN 6 classifiers. This kind of interpretation is possible due to the fact that we use RBF networks [3], which do an interpolation between samples of a certain feature space region.

An advantage of the hierarchical classification scheme is the refinement of the input space to 'parts of interest' that is e.g. the glaucomatous perimetries. ANN 1 through ANN 6 were trained by 2/3 and tested/evaluated by 1/3 of the perimetry-sample sets according to the classification tree. Table 1 gives an overview of the sensitivity/specificity evaluation results of the ANN classifiers.

Hierarchy   Sensitivity / Specificity
ANN 1 83 % / 81 %
ANN 2 85 % / 93 %
ANN 3 91 % / 92 %
ANN 4 72 % / 71 %
ANN 5 79 % / 56 %
ANN 6 62 % / 74 %

Table 1: Results of the specialised ANN

Fuzzy Rule Sets

The overall decision about the differential diagnosis is given by a fuzzy rule set. The main structure is shown in Figure 2. CDR, IOP and the left-right-difference values are input directly as real numbers which will be fuzzified by membership functions (MBF ) defined by the medical expert in co-operation with the knowledge engineer. The first rule level combines these values using min-max decision rules to get output values of the first level : These are intraocular hypertension, normotensive glaucoma and all other glaucoma types (glaucoma). The result is given as a membership value to the variable terms 'yes' and 'no'. These provisional results are used as an output and input to the next rule level, respectively. The second rule level gets additional input from the classification of perimetry data. As shown in Figure 2 this preclassification step can be done by the medical expert or automatically by the neural network classifier. The latter is our solution.

Main structure fuzzy rule set Figure 2: Main structure fuzzy rule set(large)

The classification of the perimetry data is input at all three hierarchical levels. Starting with the highest one (normal-pathological-not acceptable) down to the deeper classification of normal and pathological/glaucomatous classes. These ideas are realised using the software package fuzzyTECH. The final decision outcome (final_flaucoma_yes etc.) is combined to 'situation classes' like 'glaucomatous changes', 'suspect glaucomatous changes', 'pathological but seemingly not glaucomatous changes', 'normal'. This final decision is presented to the user as a verbal message with a short description and the possibility to get further information sending a query to the ophthalmic knowledge-based information system [4].


This combination of automatic classification by hierarchical ordered neural nets and fuzzy rule sets at different levels was applied to a test data set. It was collected during the last ophthalmologist's consultation of 30 patients already known in their history to the monitor. The classification supporting the differential diagnostic decision gave acceptable results to the medical user because of the fuzzy strategy and a trend to 'summarise' even small membership values for each single perimetry class in connection with the degree of support of the rules [5]. Further developments include testing of the above described approach with larger data sets and the implementation of time dependencies of the development among situation classes.

Acknowledgements: This work is part of the OPHTEL project funded by the European Commission within the 'Telematics Application Programme'.


[1] S.E. Spenceley, D.B. Henson, D.R. Bull, Visual field analysis using artificial neural networks, Ophthal. Physiol. Opt. 14 (1994) 239-248.
[2] A. Newell, The Knowledge level, Artificial Intelligence 18(1982), 87-127
[3] J. Moody and C. Darken, Learning with localized receptive fields, in Proceedings of the 1988 Connectionist Models Summer School, G. Hinton, T.Sejnowski, and D. Touretzsky, eds, 133-134
[4] G.Zahlmann, M. Schubert, M. Obermaier, G.Mann,Concept of a knowledge-based monitoring system for glaucoma and diabetic retinopathy using a telemedicine approach, Proc. of the 13th Annual International Conf. IEEE-EMBS, IEEE Press 1996, 111pp
[5] G. Zahlmann, M. Scherf, A. Wegner, A neuro-fuzzy-classifier for a knowledge-based glaucoma monitor, AIME 1997
[6] M. Scherf, A New Approach to Feature Selection, AIME 1997