Applications of Deep Learning and Artificial Intelligence in Retina

The potential application of artificial intelligence for diagnosis and management of glaucoma in adults

“…In contrast, unsupervised learning is where raw input data is processed by the algorithm and divided into groups, which may or may not match the existing clinical knowledge. The term 'black box' is used in reference to deep learning algorithms given that the criteria used to make the diagnosis are unknown 14,15 . Increasingly recent studies are using hybrid methods, combining both machine and deep learning algorithms, as seen in Tables 1, 2 and 3.…”

“…Deep learning computational processes do not adhere to set guidelines but instead are developed by the computer through pattern recognition through thousands of training images such as the trials of Ting et al 25 , Li et al 29 and Medeiros et al 36 which used data inputs in excess of 32,000 images. Although deep learning algorithms have proven to be highly sensitive and specific, it is possible that computers may incorporate non-retinal related features such as artefact 59 , poor pupillary dilation or the presence of a media opacity into their analyses which may possibly confound the results 14 . Some concern has been raised by physicians and patients that the 'black box' paradigm may leave us in the dark as to how the algorithm has reached its results 58 , i.e.…”

“…It can however be argued, that deep learning could be a valuable training tool for junior doctors and an adjunct for more challenging cases where there is diagnostic uncertainty or where a second opinion is desirable. Deep learning may also help reduce human error 14 thereby raising consistency across medical professionals. Over reliance on technology may potentially be harmful at times if/when technology fails, and "output" is accepted without question.…”

“…However, infinitely increasing the size of the dataset used may not necessarily improve the diagnostic performance of the algorithm and instead may increase the risk of false connections forming 10 . As highlighted above, it is possible that the algorithms are incorporating non-retinal related features 14,59 .…”

“…AI and deep learning techniques offer a tantalising promise of more precise and earlier detection of sight-threatening disease. This would focus the attention of both patient and ophthalmologist on the importance of compliance with treatment and maintaining follow up 14 . Earlier detection and more intensive, personally-targeted treatment of glaucoma may help slow or arrest the disease progression and allow patients to maintain their independence, their career and driving licence for longer.…”