|
|
COMMENTARY |
|
Year : 2019 | Volume
: 67
| Issue : 7 | Page : 1011 |
|
Commentary: AI for an eye leaves the whole world…
Ashwin Mohan, Rashmin Gandhi
Beyond Eye Care, Mumbai, Maharashtra, India
Date of Web Publication | 25-Jun-2019 |
Correspondence Address: Dr. Ashwin Mohan Beyond Eye Care, Mumbai, Maharashtra India
Source of Support: None, Conflict of Interest: None | Check |
DOI: 10.4103/ijo.IJO_624_19
How to cite this article: Mohan A, Gandhi R. Commentary: AI for an eye leaves the whole world…. Indian J Ophthalmol 2019;67:1011 |
Ophthalmology is a branch of medicine that deals predominantly with history and visual examination. Auditory clues (bruits) or palpation and percussions (pulsations) are very few in number. This means that all the information needed to diagnose the majority of the ophthalmic conditions can be fed into a computer very easily. This information is the raw material on which artificial intelligence (AI) algorithms can then process an output. The key to a good output is good input, and hence, this is an advantage for AI in ophthalmology.
An article in this issue of IJO reviews the current literature for AI in diabetic retinopathy (DR).[1] This is an apt topic to enter the vast world of AI in ophthalmology as DR is not only a very prevalent disease but also has a number of visual examination findings (microaneurysms, dot and blot hemorrhages, hard exudates, flame-shaped hemorrhages, cotton wool spots, vascular alterations, membranes, and neovascular alterations), which can be taught to a computer. The main characteristics that are taught to a computer include the color, shape, location, and contrast from the background. This and annotations by a trained expert that include marking the exact pathologies on the image is what constitutes the “ground truth” for an algorithm. This is the foundational base and thus of immense importance. It is on this base that machine learning can then fine-tune the output.
For most softwares this ground truth has been aquired from a team of doctors and not from a single individual. The technical jargon and philosophy are completely different in the medical and engineering field, and hence, there is room for miscommunication. Lack of communication between the doctors and the engineers coding the software is thus one of the pitfalls; heterogeneity and inaccuracy of information are the others. Because the entire ground truth has not been obtained from a single individual, there is room for conflicting entries, and thus, a cause for future “bugs” in the system. However, if the entire ground truth is obtained from a single individual there is a chance of it being completely wrong. Hence, the team of doctors needs to be chosen wisely so as to achieve a balance between accuracy and heterogeneity.
Speaking of balance another important parameter that can be titrated is the sensitivity and specificity of the software. This is usually done by setting a specific cut-off for each of the parameters. As a general rule, specificity is inversely proportional to sensitivity – hence, if you do not want to miss out on identifying any subject with disease (high sensitivity), you have the risk of falsely identifying a few (more false positives hence less specific) or vice versa. At first sight, this may seem completely alien to how we humans think, however, it has more similarities than what meets the eye; no pun intended. We as doctors are taught to have a “high index” of suspicion when there certain risk factors present for a few diseases, which in effect translates to increase our sensitivity and lowering our specificity. We must understand that AI is nothing but a simulation of RI or real intelligence - our collective human conscience. Hence, these algorithms can be programmed to adjust their thresholds depending on a few patient parameters such as elder age, duration of disease, sugar levels, etc.
However, we come back to one question – do we need AI? Do we need AI when we have RI? Most of the reasoning behind AI revolves around how it is automated, and hence, can reach areas with limited access to healthcare – i.e. reach where RI cannot, and hence, it is not a replacement for humans but works as an assistant. This is an essential ego boost critical to the growth of AI in a world where a large number of doctors feel threatened by the astronomical growth of AI applications in the last decade. The fact that AI works unaffected by emotions is an aspect not much talked about. The emotional state of a human has significant effects on the decision-making and hence leads to variability (inter and intraobserver) in the clinical decisions. AI offers a more stable state and hence can be more predictable.
Thus, we have seen that a good ground truth is the foundation of AI, adequacy of data input is an advantage for AI in ophthalmology, tweaking thresholds depending on risk factors can make AI more dynamic and relevant, AI can be more consistent and predictable than the human mind, and can reach places where humans cannot; thus, leaving us human doctors to do the most “human” thing in healthcare, something conspicuously absent from many of us – the human touch. It will be exciting to see if in future AI can help make healthcare more predictable and reliable, whereas we doctors make it more human and compassionate.
References | | |
1. | Padhy SK, Takkar B, Chawla R, Kumar A. Artificial intelligence in diabetic retinopathy: A natural step to the future. Indian J Ophthalmol 2019;67:1004-9. [Full text] |
|