A Deep Learning Model for Glaucoma Detection Outperforms Conventional Metrics.

PURPOSE

To test a deep learning model (DLM) for detecting glaucomatous damage with optical coherence tomography (OCT) reports.

METHODS

An ImageNet-pretrained ResNet-50 model was fine-tuned to classify healthy (H) versus glaucomatous (G) OCT reports. The model was trained with 4932 H and 207 G OCT wide-field reports and tested on three unseen datasets. The "Clear G" cohort consisted of 50 eyes with clear glaucomatous damage based on OCT, visual fields, photos, and clinical exams. The "Pattern ON-G" cohort contained 183 eyes over 60 years of age with OCT arcuate defects. These defects ranged from subtle to advanced. The H-test cohort contained 396 eyes from a commercial normative reference database. We compared the DLM to five conventional metrics.

RESULTS

For a DLM cutoff yielding a 99.5% specificity on the 396 H-test eyes, the DLM achieved a sensitivity of 100% with the 50 Clear G and 95.1% with the 183 Pattern ON-G reports. The AUROC was 0.999. The DLM was significantly better than 3 conventional metrics. For example, global circumpapillary retinal nerve fiber thickness showed a specificity of 97.7% for the 396 H eyes, sensitivities of 73.8% and 44.1% for the G cohorts, and an area under the receiver operating characteristic curve (AUROC) of 0.883. The best of the conventional metrics, which was based on a logistic regression model, also had significantly poorer sensitivities (84.6% and 66.1%) for the G cohorts and a lower AUROC (0.975).

CONCLUSIONS

The DLM achieved high specificity and sensitivity and outperformed conventional metrics.

TRANSLATIONAL RELEVANCE

The DLM provides a possible method for screening, and a possible diagnostic aid for clinicians.