Purchase this article with an account.
Wenji Wang, Haogang Zhu, Ya Xing Wang, David Garway-Heath; A transparent deep learning method for diagnosis of early glaucoma with optical coherence tomography. Invest. Ophthalmol. Vis. Sci. 2017;58(8):4000.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a clinically understandable deep neural network for diagnosis of early glaucoma. Aside from a single probability score output from a convolutional neural network(CNN), local defects are also inferred to enable transparent diagnosis.
The parapapillary retinal nerve fiber layer thickness(RNFLT) measured by Spectralis OCT(Heidelberg Engineering) were extracted from an elderly Chinese population with mean age of 64.6 ± 9.8 years (Beijing Eye Study). In 6046 eyes from 3316 subjects, 478 eyes (359 subjects) were diagnosed as glaucoma and 5568 eyes (2957 subjects) were healthy controls in which any retinal or optic neurological diseases other than glaucoma were excluded. A 5-layer deep CNN (Fig-a) was constructed to classify each RNFLT into glaucoma and non-glaucoma. Seventy-five percent of the data were randomly selected for training and the remaining 25% were used for validation. Furthermore, to quantify the local defect, a reference was inferred for each individual RNFLT by gradient decent with respect to the measured RNFLT such that the measured one is gradually deformed into its ‘healthy’ status. The area under the reference RNFLT(rRNFLT) was summarised as the diagnosis criteria, which is clinically understandable. ROC analysis was used to compare the CNN probability score, area under rRNFLT and a conventional support vector machine(SVM) classifier.
The inference of the rRNFLT for each individual converges on every case. The mean (standard deviation, std) area under rRNFLT is 40.16 (6.53)[µm×degree] and 47.70 (4.95)[µm×degree] for glaucoma and healthy subjects. The mean (std) CNN probability score is 0.68 (0.17) and 0.32 (0.17) for the two groups respectively. The area under ROC curve is 0.851, 0.828, 0.778 for CNN probability score, area under rRNFLT and SVM. At 15% false-positive rate, the sensitivity is 0.69, 0.62, 0.47 for the three methods under comparison(Fig-b).
The CNN model extracts comprehensive features of training samples, and it detects local defects on retinal never fiber layer(Fig-c). There is no significant difference between CNN probability score and area under rRNFLT and these two methods are significantly better than SVM. The area under rRNFLT is a transparent method providing clinically understandable information for the diagnosis of glaucoma. The method can also be extended to other diseases as well as other forms of measurements.
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