Table 1.
Demographic characteristics of the training group.
Fig 1.
The shape of the tensor (input/output) is described on the right side of each layer box. The global average pooling layer and four fully connected network (dense) layers were connected after the Inception V3 backbone CNN (convolutional neural network) architecture. The four dense layers used ReLu (rectified linear unit) as the activation function.
Fig 2.
Visual field test pattern and Garway-Heath map.
(A) A colour fundus photo with Humphrey 24–2 visual field test pattern and two optical coherence tomography (OCT) images were overlapped on the fundus photo. Garway-Heath sectorisation (white radiating line) is drawn on the optic nerve head centre. (B) Regions of visual field test points outlined by Garway-Heath sectorisation map. The central dashed square shows the boundary of the macular OCT scan area and the surrounding area is defined as the peripheral OCT scan area. IN: inferonasal, IT: inferotemporal, N: nasal, SN: superonasal, ST: superotemporal, T: temporal.
Table 2.
Demographic characteristics of the test group.
Fig 3.
Representative cases of visual field prediction.
(A) The combined OCT images, which were input into the deep learning architecture, are shown on the left column. The actual threshold values of visual field exams are shown in the (B) middle panel and the threshold values predicted by Inception V3 based deep learning architecture are shown on the (C) right panel. The colour reference for the threshold values are shown at the bottom. Despite the artificial intelligence having never seen the actual visual field, the predicted visual field looked very similar to the actual visual field exam.
Table 3.
Global and regional root mean square error of visual field prediction.
Fig 4.
Representative example of Class Activation Map (CAM).
The figure shows (A) the actual threshold values and (B) the predicted threshold values of the visual field examination. (C) Fifty-two CAMs were placed at individual visual field test points. Each CAM image is numbered at the top left. (D) Structure-function mapping between the combined OCT image (left) and the visual field (right). The macular scan in the combined OCT image corresponds to the dashed rectangle in the visual field. Color-coded Garway-Heath sectors are superimposed on the ONH scan of the combined OCT image and the corresponding visual field regions are similarly colored. The numbers in the visual field image are the same as those in the CAM images of (C).
Table 4.
Correlation coefficients and simple linear regression analyses between visual field prediction error and various factors.
Table 5.
Multiple linear regression analyses between visual field prediction error and various factors.
Fig 5.
Scatter plot of the prediction error ratio (macular/peripheral) versus the visual field Mean Deviation (MD).
The slope was –0.020 (P < 0.001) which suggested that as the MD decreased, the macular prediction error became higher than the peripheral prediction error. In other words, as glaucoma progressed, the peripheral prediction became more accurate than the macular prediction.