Purchase this article with an account.
Kaveri A Thakoor, Qian Zheng, Linyong Nan, Xinhui Li, Emmanouil (Manos) Tsamis, Rashmi Rajshekhar, Isht Dwivedi, Iddo Drori, Paul Sajda, Donald C Hood; Assessing the Ability of Convolutional Neural Networks to Detect Glaucoma from OCT Probability Maps. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1464.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the performance of convolutional neural network (CNN) architectures, with and without previous training on non-medical images, at distinguishing between healthy controls and glaucoma suspects using retinal nerve fiber layer (RNFL) probability maps from wide-field optical coherence tomography (OCT) imaging.
CNN architectures were evaluated on data from 343 glaucoma or suspect eyes (patients). CNN A consisted of 9 layers (3 convolutional blocks, each composed of a two-dimensional convolutional layer with 32 filters and 3x3 kernel size followed by Rectifying Linear Unit (ReLU) activation function followed by maximum pooling). CNN B consisted of the ResNet-18 architecture  pretrained on ImageNet . Both networks were followed by Random Forest classifiers. CNN A was trained on 215 healthy eyes (controls) and 193 patients. Both models were validated on 100 control and 50 patient eyes and tested on 100 control and 100 patient eyes. All eyes had RNFL probability maps derived from wide-field OCT images (Topcon). An OCT expert rated each eye, further dividing the eyes into 192 definitely glaucoma patients, with the remaining eyes added as controls. Performance for CNNs A and B was evaluated on this dataset (with number of training, validation, and testing samples held comparable to that used for the original 343-patient dataset). False positives (FP) and false negatives (FN) were analyzed.
On the 343-patient dataset, CNN B’s validation, 97.3%, and testing, 95.3%, accuracies were comparable to those, 96.6% and 96.9%, of CNN A. On the newly-rated dataset of 192 glaucoma patients, the validation and testing accuracies of CNNs A and B remained similar: 96.6% (validation) and 97.0% (testing) for CNN A, 95.9% (validation) and 94.4% (testing) for CNN B. FP and FN will inform future training.
CNNs A and B distinguished glaucoma patients from controls with high accuracy, moving toward automated glaucoma detection in places/times when human experts may not be accessible. A large amount of training data is not needed if CNNs previously trained on images from a different modality (such as ResNet-18) are harnessed, followed by feature extraction and classification using non-parametric techniques (such as Random Forest classifiers), enabling a transfer learning bridge between natural images and medical images.  He et al., CVPR 2016.  Russakovsky et al., IJCV 2015
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only