The field of artificial intelligence has undergone a resurgence in recent years, primarily because of the significant advancements in image recognition through deep learning (Figure 5). Artificial intelligence is being increasingly implemented in medical fields that rely heavily on imaging, including the field of retina. Success has been achieved in image segmentation and classification, and computer-aided diagnosis models are being approved by regulatory bodies with FDA approval for the first medical device using artificial intelligence granted in 2018.[69–73] After segmentation and diagnosis, the next frontier of artificial intelligence in the field of retina is in predictive modeling which is being developed to predict disease progression and treatment response[74–77] including visual acuity prediction after receiving injections for AMD.
Artificial intelligence models have also illuminated clinical correlations that were previously unimaginable. Deep learning algorithms have predicted demographics, cardiovascular risk factors, and anemia from fundus photos alone,[79,80] and future algorithms may provide increasing associations between neurodegenerative and cardiovascular disorders.[81,82] Artificial intelligence systems perform at such a high level through integrating large volumes of data to find subtle patterns among millions of pixels in fundus photographs and billions of voxels in three-dimensional optical coherence tomography scans.
However, many of these models have not been validated on large external real-world datasets. As we are currently tied to methods which require large volumes of curated and labeled training data, the setting of ground truth for these input images has a considerable impact on the final performance metrics of the resulting artificial intelligence models. It becomes increasingly important to set strict labeling and adjudication criteria for these image labels. Diagnostic accuracy studies comparing physicians, artificial intelligence models, and artificial intelligence-augmented physicians are necessary to determine the net benefit of this technology. A recent systematic review was the first of its kind in comparing performance between providers and deep learning for detecting disease in medical imaging. It found artificial intelligence model performance to be equivalent to providers; however, few of the articles analyzed reported diagnostic accuracy with externally validated results.
The retina specialist should not be undervalued, as effective retinal medicine will benefit from artificial intelligence enhancement and not by artificial intelligence replacement. Artificial intelligence will lift and homogenize accuracy of retinal diagnosis while enabling personalized care by predicting functional outcomes and treatment response. This integration of data will provide more time with patients, whether to determine a treatment plan or to contextualize the functional impact of a patient's predicted visual changes. The new high-value skills will include interpreting and personalizing recommendations made in conjunction with artificial intelligence, and as more time avails, retina specialists will have more 'time to be human' with their patients.
Curr Opin Ophthalmol. 2020;31(3):207-214. © 2020 Lippincott Williams & Wilkins