Abstract
Purpose: :
To compare accuracy and classification error scores of visual field assessment performed by clinicians to that performed by an artificial neural network (ANN).
Methods: :
We have earlier described a fully trained ANN with high diagnostic performance, using probability scores from pattern deviation probability maps as input data. We compared the performance of our ANN to that of thirty clinicians with different experience of glaucoma management. Glaucoma patients were selected in pseudo random order from one of our perimeters. Eligibility required a diagnosis of glaucoma and a glaucomatous optic disc, but visual field defects were not necessary. Only eyes with MD better than -10dB were included. Fields from healthy subjects were randomly sampled from an existing database. 30-2 SITA Standard visual field printouts with full Statpac information from 99 glaucoma patients and 66 healthy subjects were independently assessed in a masked fashion. A grading scale from 1 to 10 was used for assessment, where 1 indicated absolute certain healthy, 10 absolute certainty of glaucoma; 5.5 was the cut-off between healthy and glaucoma. To match the grading scale, the continuous logistic function output of the ANN was transformed into a linear scale and divided into 10 equal intervals. An error score was calculated for each field classification by counting the number of steps from the correct endpoint, i.e. from 1 for healthy subjects and from 10 for glaucoma patients, to the actual score. The mean classification error score for each clinician and the ANN was then calculated as the average of the error score for each field.
Results: :
Sensitivity of the ANN was 93%, which was significantly better (p<0.001, McNemar’s test) than the average of the clinicians 83%. Specificities were similar, 91% for the ANN and 90% for clinicians. Mean classification error score of the ANN (1.50) was amongst the best third of the clinicians (mean of 1.80), and also nearly as good as that of the glaucoma experts (1.48). The ANN, as opposed to the clinicians, never indicated a high degree of diagnostic certainty for any misclassified visual field test.
Conclusions: :
Our trained ANN classified visual fields for glaucoma at least as well as clinicians.
Keywords: clinical (human) or epidemiologic studies: systems/equipment/techniques • visual fields • imaging/image analysis: clinical