Purchase this article with an account.
Bhavna Josephine Antony, Stefan Maetschke, Hiroshi Ishikawa, Gadi Wollstein, Joel S Schuman, Simon Wail; Feature agnostic networks outperform classical machine learning approaches in the detection of glaucoma in OCT volumes. Invest. Ophthalmol. Vis. Sci. 2019;60(9):2206. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To test the performance of a feature agnostic deep learning approach that labels raw OCT image data (macular and optic nerve head (ONH)) as healthy or glaucomatous in comparison with classical machine learning approaches.
OCT scans were acquired from both eyes of 134 healthy, 779 glaucoma patients (Cirrus HD-OCT scanner, 200x200 Macular and Optic Disc Cubes, Zeiss, Dublin CA). Two convolutional neural network (CNN) were trained to label the OCT volumes (ONH and macular) as healthy or glaucomatous. The cubes were downsampled to 64x64x96 voxels. In a 5-fold cross validation experiment, the model was trained using cross entropy as the loss function and area under the receiver operating characteristic curve (AUC) as the performance metric. The OCT volumes were not denoised, segmented or corrected for laterality. The CNN performances were also compared to classical machine learning approaches - logistic regression (LR) and random forests (RF) trained with the following macular and ONH features separately. The macular features: retinal nerve fibre layer (RNFL) and ganglion cell-inner plexiform complex (GCIPL) thickness (global mean and 6 oval regions), and total macular thickness in the 9 ETDRS regions. The ONH features: circumpapillary RNFL thickness (global mean, 12 clock hours and 4 quadrants), cup-to-disc ratio (vertical and mean), and the cup volume. Class activation maps (CAM) were also computed to visualise the regions identified by the network.
In the ONH scans, LR showed the best performance among the classical machine learning methods with a mean (SD) AUC = 0.89 (0.03). The CNN outperformed this technique which was signficantlyhigher AUC = 0.94 (0.04) (p < 0.01). The RF showed the best performance on the macular features with AUC = 0.66 (0.11), which was again outperformed by the CNNs with an AUC = 0.85 (0.01) (p < 0.01). CAMs in the ONH volumes highlighted the cup and retinal region along the arcades, while the inner retina (excluding the RNFL) were highlighted in the macular volumes (see Fig. 1)
The feature agnostic approaches greatly outperformed classical techniques that rely on segmented features, suggesting their robustness. Since feature agnostic approaches do not require any pre-processing or segmentation of structures, potential impact of segmentation failure can be eliminated.
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