Biomedical subjects

S E Spenceley

Publications and source records attributed to S E Spenceley.

4 recordsLinked to original sources

PubMed1997

Artificial neural network analysis of noisy visual field data in glaucoma.

This paper reports on the application of an artificial neural network to the clinical analysis of ophthalmological data. In particular a 2-dimensional Kohonen self-organising feature map (SOM) is used to analyse visual field data from glaucoma patients. Importantly, the paper addresses the problem of how the SOM can be utilised to accommodate the noise within the data. This is a particularly important problem within longitudinal assessment, where detecting significant change is the crux of the problem in clinical diagnosis. Data from 737 glaucomatous visual field records (Humphrey Visual Field Analyzer, program 24-2) are used to train a SOM with 25 nodes organised on a square grid. The SOM clusters the data organising the output map such that fields with early and advanced loss are at extreme positions, with a continuum of change in place and extent of loss represented by the intervening nodes. For each SOM node 100 variants, generated by a computer simulation modelling the variability that might be expected in a glaucomatous eye, are also classified by the network to establish the extent of noise upon classification. Field change is then measured with respect to classification of a subsequent field, outside the area defined by the original field and its variants. The significant contribution of this paper is that the spatial analysis of the field data, which is provided by the SOM, has been augmented with noise analysis enhancing the visual representation of longitudinal data and enabling quantification of significant class change.

Diagnosis, Computer-Assisted↗

PubMed1996

Visual field test simulation and error in threshold estimation.

AIM: To establish, via computer simulation, the effects of patient response variability and staircase starting level upon the accuracy and repeatability of static full threshold visual field tests. METHOD: Patient response variability, defined by the standard deviation of the frequency of seeing versus stimulus intensity curve, is varied from 0.5 to 20 dB (in steps of 0.5 dB) with staircase starting levels ranging from 30 dB below to 30 dB above the patient's threshold (in steps of 10 dB). Fifty two threshold estimates are derived for each condition and the error of each estimate calculated (difference between the true threshold and the threshold estimate derived from the staircase procedure). The mean and standard deviation of the errors are then determined for each condition. The results from a simulated quadrantic defect (response variability set to typical values for a patient with glaucoma) are presented using two different algorithms. The first corresponds with that normally used when performing a full threshold examination while the second uses results from an earlier simulated full threshold examination for the staircase starting values. RESULTS: The mean error in threshold estimates was found to be biased towards the staircase starting level. The extent of the bias was dependent upon patient response variability. The standard deviation of the error increased both with response variability and staircase starting level. With the routinely used full threshold strategy the quadrantic defect was found to have a large mean error in estimated threshold values and an increase in the standard deviation of the error along the edge of the defect. When results from an earlier full threshold test are used as staircase starting values this error and increased standard deviation largely disappeared. CONCLUSION: The staircase procedure widely used in threshold perimetry increased the error and the variability of threshold estimates along the edges of defects. Using earlier data, when available, overcomes this problem and reduces examination time.

Computer Simulation↗

PubMed1996

Spatial classification of glaucomatous visual field loss.

AIMS: To develop and describe an objective classification system for the spatial patterns of visual field loss found in glaucoma. METHODS: The 560 Humphrey visual field analyser (program 24-2) records were used to train an artificial neural network (ANN). The type of network used, a Kohonen self organising feature map (SOM), was configured to organise the visual field defects into 25 classes of superior visual field loss and 25 classes of inferior visual field loss. Each group of 25 classes was arranged in a 5 by 5 map. RESULTS: The SOM successfully classified the defects on the basis of the patterns of loss. The maps show a continuum of change as one moves across them with early loss at one corner and advanced loss at the opposite corner. CONCLUSIONS: ANNs can classify visual field data on the basis of the pattern of loss. Once trained the ANN can be used to classify longitudinal visual field data which may prove valuable in monitoring visual field loss.

Glaucoma↗

PubMed1994

Visual field analysis using artificial neural networks.

There have been several reports on the application of artificial neural networks (ANNs) to visual field classification. While these have demonstrated that neural networks can be used with good results they have not explored the effects that the training set can have upon network performance nor emphasized the unique value of ANNs in visual field analysis. This paper considers the problem of differentiating normal and glaucomatous visual fields and explores different training set characteristics using field data collected from a Henson CFS2000 perimeter. Training set properties including size, balance between normals and glaucomas, extent of field loss and the spatial location of glaucomatous defects are explored. A multilayer network with 132 input nodes, 20 hidden layer nodes and 2 output nodes in trained using an error backpropagation algorithm. Both sensitivity and specificity are measured during testing. The results demonstrate that large random sets are better than small random sets since sensitivity improves with size and specificity is not adversely affected. The variability in performance also reduces as training set size increases. In addition, sets that are biased towards glaucoma examples are more sensitive and less specific, while sets biased with normal examples are more specific and less sensitive than balanced sets. Thus large training sets with class balance are generally desirable for good sensitivities and specificities. The actual glaucoma examples contained in the set are also important. A training set deficient in examples has no detrimental effect on sensitivity or specificity. The spatial distribution of defects is also crucial. Spatially biased sets are unable to recognize defects that occur in locations where no previous defect has been presented while more balanced sets lead to better performance. In conclusion the 'ideal' training set should contain many examples of early defects that represent the full range of locations where these defects may occur.

Algorithms↗