|
 
 |
| EDITORIAL |
|
| Year : 2022 | Volume
: 12
| Issue : 2 | Page : 121-122 |
|
Current status of artificial intelligence for medicine
Hitoshi Tabuchi
Department of Technology and Design Thinking for Medicine, Hiroshima University, Higashihiroshima; Department of Ophthalmology, Tsukazaki Hospital, Himeji, Japan
| Date of Submission | 14-Apr-2022 |
| Date of Acceptance | 14-Apr-2022 |
| Date of Web Publication | 01-Jun-2022 |
Correspondence Address:
Dr. Hitoshi Tabuchi
Department of Technology and Design Thinking for Medicine, Hiroshima University, 1-2-3 Kasumi, Minami-Ku, Hiroshima, 734-8553
Japan
 Source of Support: None, Conflict of Interest: None
DOI: 10.4103/tjo.tjo_23_22
How to cite this article:
Tabuchi H. Current status of artificial intelligence for medicine. Taiwan J Ophthalmol 2022;12:121-2 |
The application of deep learning in ophthalmology began with the diagnosis of diabetic retinopathy by a Google team,[1] and they are currently actively working on its real-world implementation in Thailand.[2] My team has also reported a wide range of artificial intelligence (AI)-based fundus diagnostic applications for patients in Japan, starting with retinal detachment diagnostic application using an ultra-widefield fundus camera.[3] As we accumulate our experiences in this area of study, we have come to understand the fundamental limits of deep learning, that is based on supervised learning. The performance of the AI model is highly dependent on the training data set. The face recognition system developed using mainly facial images of lighter males had a low recognition performance of <70% on recognizing darker females; this research finding shocked the world. The number of images of darker females was not sufficient in the training data in the first place.[4] There are several global companies that are developing their own face recognition systems in Japan, where I live. It is not hard to imagine that a system that works in Japan, which is known for its monoethnicity, does not necessarily work in other nations that are rich in diversity. The problem that our team encountered early on in our AI studies for fundus diseases was tigroid fundus. The distinct tessellated pattern is a more prominent feature for nonprofessionals than a small bleeding spot, and we wondered whether or not AI would be able to distinguish between the two. In fact, there are few patients with diabetic retinopathy who have myopic fundus. On the other hand, retinal vein branch occlusion has been reported to have an association with myopia; its association with glaucoma is well known. In the early stage of our study, we reported at an academic conference in Japan that AI distinguishes the normal fundus from the disease-free tigroid fundus.[5] Many scientific studies reported that the distribution of refraction values differs by race, which means that there are racial differences in fundus findings. The first thing that AI study researchers need to decide is framing. They have to define who would benefit from their AI application. On the other hand, what is required in product development is great marketability. If the users will be limited and the marketability would be small, the development of an application often does not start in the first place. With this point of view, diabetic retinopathy, which is not easily affected by the myopic fundus, may be a good target for AI applications because of its large framing and wide marketability. I would like to take this opportunity to present an important perspective on the real-world implementation of AI applications by giving an example of special circumstances in Japan. The prior probability of many diseases in Japan is low. The prevalence of many eye diseases, other than glaucoma and cataract, is very low in Japan. Diabetic retinopathy, one of the most predominant eye diseases in the world, is not so in our country. In addition, because ophthalmologists are available in most parts of Japan, except for some remote islands and mountainous areas, retinopathy screening is being conducted sufficiently. This means that, in Japan, situations in which diabetic retinopathy AI could be fully utilized are limited.
Although AI development is being driven by global companies, medical applications with infinite performance that would work effectively in any area or situation are still yet to be developed. In other words, current AI research and development is still in the early days; all kinds of ideas can be accepted, and it is an area that continues to offer opportunities to young people who are particularly hopeful. In addition to this editorial review that summarizes important perspectives when considering AI applications, we have presented our efforts in providing an answer to the question of whether surgical evaluation technology using AI can be used for cataract surgical training. Although this preliminary study does not provide definitive answers to the question, I hope that the study illustrated that there are still many possible applications of AI to various clinical areas, not limited to fundus diagnosis.
My AI team is characterized by a wide age group, ranging from engineers in their 60s who have retired from technology companies to medical students in their teens. We are enjoying developing many AI applications. The outcomes of these AI applications themselves may not be achieved yet, but our members' days spent researching are enjoyable and fulfilling. I personally think that the outcome of AI research at this point is the development of human resources, which will play an active role in the development of AI applications for the world in the future.
| References | |  |
| 1. | Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 2016;316:2402-10.  |
| 2. | Beede E, Baylor E, Hersch F, Iurchenko A, Wilcox L, Ruamviboonsuk P, et al. A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. Available from: https://www.dl.acm.org/doi/abs/100.1145/3313831.3376718. [Last acessed on 2021 Oct 17]. |
| 3. | Ohsugi H, Tabuchi H, Enno H, Ishitobi N. Accuracy of deep learning, a machine-learning technology, using ultra-wide-field fundus ophthalmoscopy for detecting rhegmatogenous retinal detachment. Sci Rep 2017;7:9425. |
| 4. | Buolamwini JA. Gender Shades: Intersectional Phenotypic and Demographic Evaluation of Face Datasets and Gender Classifiers (PhD Thesis). MIT; 2017. Available from: https://www.hdl:1721.1/114068.OCLC1026503582. [Last accessed on 2021 Oct 17]. |
| 5. | Masumoto H, Tabuchi H, Ohsugi H, Ishitobi N. Evaluation of Tigroid Fundus Discrimination Using Deep Neural Network. Japan Myopia Society Annual Meeting; 2018. |
|
Search Pubmed for