Using Deep Learning to Automate Goldmann Applanation Tonometry (GAT)

Abstract

Purpose : Traditional GAT techniques are susceptible to user bias. We developed an objective and automated method for measuring IOP using deep learning on videos captured with standard GAT measuring techniques.

Methods : Subjects were recruited from an academic glaucoma clinic. For automated GAT, an iPod Touch (Apple, Cupertino, CA) was clamped on a standard slit lamp microscope and the Goldmann tonometer was set to a fixed force (Fig 1a). A video of the applanation mires generated with this fixed force was acquired. Standard GAT was also performed. Videos were split into training and validation videos in an 80:20 ratio. Frames from training videos (Fig 1b) were labeled to identify the reference tonometer diameter and the outline of the fixed force applanation mires (Fig 1c). Labeled images were used to train a deep learning model to automate mire localization. IOP values were calculated from the mire diameters using the Imbert-Fick formula. A separate prospective test set was collected where two independent masked observers measured standard and automated GAT to assess inter-observer and test-retest variability.

Results : 263 eyes from 135 subjects were included in the training and validation videos. For the test set, 50 eyes from 25 subjects were included, which generated 100 videos as each eye was measured by two observers. Within the test set, the mean difference between automated and standard GAT was -0.9 mmHg with 95% limits of agreement (LoA) of -5.4 to 3.6 (Figure 2a). Mean difference between the 2 observers using standard GAT was 0.09 mmHg with LoA of -3.8 to 4.0 (Figure 2b). Mean difference between the 2 observers using automated GAT videos was -0.3 mmHg with LoA of -4.1 to 3.5 (Figure 2c). The mean absolute interobserver differences did not differ significantly between standard and automated GAT (1.5±1.3 versus 1.4±1.4 mmHg, respectively, P=0.6). When IOP measurements for entire videos were plotted, ocular pulsations could be observed (Fig 1d).

Conclusions : Preliminary measurements using deep learning to automate GAT demonstrate comparable results to standard GAT.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1. (A) Example of iPod Touch set up (B) Example of unlabeled and (C) Labeled frame with the tonometer tip and 2 mires outlined (D) Example of ocular pulsations captured by plotting IOP versus time

 

Figure 2. Bland-Altman plots comparing IOP values from: (A) Standard vs automated GAT (B) Standard GAT by 2 observers (C) Automated GAT videos by 2 observers