Purchase this article with an account.
M. K. Smolek, J. K. Hovis; Comparison of Traditional Scoring and Neural Network Analysis of Farnsworth-Munsell 100 Hue Test Data. Invest. Ophthalmol. Vis. Sci. 2007;48(13):3812.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To design a neural network (NN) to automatically screen red-green color vision abnormalities from normal cases using data acquired with the Farnsworth-Munsell 100 Hue Test (FM-100).
120 subjects were recruited for this study; 42 females (f) with a mean age of 28.2 (range: 19 to 50) and 78 males (m) with a mean age of 32.8 (range 18 to 68). Subjects were categorized by a clinical color vision expert as either abnormal (3 f / 49 m) or normal (39 f / 29 m) based on traditional scoring of FM-100 plots. Nagel anomaloscope analysis was used as a gold standard to determine the true nature of the condition and allow comparison of FM-100 scoring to NN screening. FM-100 results for all subjects were divided randomly into training (n = 59; 25 abnormal / 34 normal) and test sets (n = 61; 27 abnormal / 34 normal) for use in backpropagation NN training. Three FM-100 values were used as inputs to the NN: Raw Total Error; Raw Red/Green Error; and Raw Blue/Yellow Error. The NN output consisted of 2 categories (normal & abnormal) with values between 0.0 and 1.0. Ten hidden neurons were used. Winner-take-all scoring >0.5 was used. A NN output value >0.5 and approaching 1.0 indicated a high likelihood of correspondence between the input pattern and the output category, while an output value <0.5 was consistent with little or no correspondence.
As shown in the table, the NN method outperformed the traditional scoring of FM-100 data with respect to sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Traditional scoring produced 7 false negatives and 3 false positives, while NN screening produced 3 false negatives and 0 false positives. One male deuteranomalous case and two male protanomalous cases were among the 3 false negatives.
The NN result using total error and partitioned error scores in a screening paradigm outperforms FM-100 traditional scoring. NNs are particularly useful as clinical screening and classification tools, particularly when an immediate result is desired for a process involving tedious data analysis or subjective interpretation of the results.
This PDF is available to Subscribers Only