|LETTER TO THE EDITOR
|Ahead of print publication
Understanding required to consider artificial intelligence applications to the field of ophthalmology
Department of Technology and Design Thinking for Medicine, Hiroshima University, Higashihiroshima, Japan
|Date of Submission||18-Jun-2022|
|Date of Acceptance||19-Jul-2022|
|Date of Web Publication||22-Sep-2022|
Department of Technology and Design Thinking for Medicine, Hiroshima University, Higashihiroshima
Source of Support: None, Conflict of Interest: None
|How to cite this URL:|
Tabuchi H. Understanding required to consider artificial intelligence applications to the field of ophthalmology. Taiwan J Ophthalmol [Epub ahead of print] [cited 2022 Nov 27]. Available from: https://www.e-tjo.org/preprintarticle.asp?id=356685
The reviewer's thought-provoking comments on our review article are appreciated. Based on these comments, information on state-of-the-art applications in artificial intelligence has been added to the text.
It is recognized that poor performance in geographically different populations (facilities) is a typical example of evaluating the data that differs from the training dataset.
There are two events where artificial intelligence (AI) can be implemented to evaluate the data from unlearned groups: Current events and future events.
- Differences in equipment (e.g., different versions of medical equipment and different manufacturers of medical equipment)
- Differences in medical personnel (e.g., the skill level of photographers and the skill level of technicians who complete staining in pathology)
- Regional differences (e.g., different races and, in the case of pathology, different pH levels of the water).
- All ongoing events
- Shifts in patient distribution (e.g., immigration, aging population)
- A shift in criteria (e.g., change in the medical criteria for determining any disease).
We can handle current events if we invest time and effort. However, future events will occur after the machine learning model is completed, and they are even more challenging to predict. Therefore, the performance of the machine learning model deteriorates over time if improvements are not made.
Machine learning operations (MLOps) are becoming increasingly important and efforts should be made in this area. From an MLOps perspective, it is essential to have a mechanism to maintain and perpetuate the performance of artificial intelligence after its implementation in society. For example, for the intraocular lens check AI that the author and his team are developing, providers must continuously replace the training data from obsolete intraocular lenses with new data.
However, as mentioned in the main text, there are objections to changes in the performance of an application certified as a medical device, for better or worse.
We would like to thank the reviewer for sharing a fascinating paper on 3D technology applications. Our team has previously published an AI application for diagnosing ERM using 3D OCT images, and we are continuing unpublished research in clinical applications combining 3D images and AI technology.
Financial support and sponsorship
Conflicts of interest
The author declares that there are no conflicts of interests of this paper.
| References|| |
Cao T, Huang CW, Yu-Tung Hui D, Cohen JP. A benchmark of medical out of distribution detection. arXiv 2007. [doi: 10.48550/arXiv. 2007.04250].
Tabuchi H, Masumoto H, Adachi S. Real-world testing of artificial intelligence system for surgical safety management. Invest Ophthalmol Vis Sci 2020;61:2032.
Sonobe T, Tabuchi H, Ohsugi H, Masumoto H, Ishitobi N, Morita S, et al
. Comparison between support vector machine and deep learning, machine-learning technologies for detecting epiretinal membrane using 3D-OCT. Int Ophthalmol 2019;39:1871-7.