Purchase this article with an account.
Kunal K Dansingani, Surya Teja Devarakonda, Kiran Vupparaboina, Soumya Jana, Jay Chhablani, K Bailey Freund, Orly Gal-Or, Sarra Gattoussi; Classification of macular lesions using optical coherence tomography and an artificial neural network. Invest. Ophthalmol. Vis. Sci. 2017;58(8):823.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Within the diverse spectrum of macular disease, most disorders have a limited number of lesion subtypes which show characteristic configurations related to the underlying diagnosis. Artificial neural networks (ANNs) can be configured as function approximators and can be trained to classify appropriately coded data. This study evaluates the training and performance of an artificial neural network in classifying optical coherence tomography (OCT) scans of a subset of lesions frequently seen in central serous chorioretinopathy (CSC), and distinguishing them from scan samples from normal eyes.
Swept-source OCT scans of healthy eyes and eyes with CSC were acquired. Scan regions containing lesions of interest were cropped, labelled and resampled to a feature vector size of 900. Experiment 1: An ANN (5 hidden layers, 2 output nodes, sigmoid activation, learning rate 0.001, 25 epochs) was trained using 70% of the data samples and tested with the remaining 30%, to distinguish normal from abnormal samples. Samples in the training data set were replicated if they had been considered, during labelling, as highly exemplary of the lesions represented. Experiment 2: A similarly configured ANN (figure) was trained to classify OCT samples as normal, or containing either pigment epithelial detachment (PED), subretinal fluid (SRF) or both.
Sample counts for training were 400,640 for normal and 21,646 for abnormal data; 21,646 samples were taken randomly from normal data for each cross-validation step. Training performance is summarized in the figure. Experiment 1: Testing yielded an accuracy of 98.6 ±0.1%, sensitivity 98.0 ±0.1% and specificity 99.3 ±0.1% for distinguishing abnormal from normal scan samples. Experiment 2: Sample counts for training were 99,928 for SRF and 19,357 for PED; 21,646 samples were taken randomly from normal data for each cross-validation step. Classification accuracy is summarised in the figure.
A sensitive and specific OCT feature classifier can be built using an ANN. Training may be accelerated by emphasising consistency in the formatting of input data and by accurate labelling of input features. The output data may be suitable for further classification to produce differential diagnostic probabilities. Acquisition of larger source data sets should reduce the need to replicate training data.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only