97th DOG Annual Meeting 1999

K22

COMPARISON OF BACKPROPAGATIONS- AND RBF - NEURAL NETS FOR THE CLASSIFICATION OF PERIMETRIC DATA SETS

G. Zahlmann, A. Wegner1, M. Scherf, M. Obermaier1

Problem: Most studies use backprogation neural nets to classify visual fields. This methods requires a 1:1 match of the input nodes with the stimulus coordinates given by the perimetric data set. Therefore we need for each perimeter a specific neural net to evaluate the visual field from this perimeter. RBF neural nets are much more flexible and allow an interpretation of every perimetry no matter from which perimeter originate. We compared sensitivity and specificity of both methods in classifiying visual fields.

Methods: 2752 visual fields (G1, Octopus 500, Interzeag) were randomly chosen and preclassified by a clinical expert into the classes 'normal'. 'glaucomatous' und 'other pathologies.

 

normal

glaucomatous

other pathologies

training

608

1192

425

validation

210

108

209



Both neural networks were trained using the first data set. For validation we used the second one unknown so far to the neural nets.

Results: The neural nets show the following sensitivities and specificities:

Neural net

Sensitivity

Specificity

RBF

72%

70%

Backpropagation

67%

78%



Conclusion: Both methods are equally efficient in classifiying visual fields.

GSF, medis, Ingolstädter Landstr. 1, D - 85764 Oberschleißheim
1Augenklinik rechts der Isar , TU München, Ismaninger Str., D - 81675 München


Back