Artificial Intelligence and Ophthalmology

Kadircan Keskinbora; Fatih Güven

doi:10.4274/tjo.galenos.2020.78989

ABSTRACT

Artificial intelligence is advancing rapidly and making its way into all areas of our lives. This review discusses developments and potential practices regarding the use of artificial intelligence in the field of ophthalmology, and the related topic of medical ethics.

Various artificial intelligence applications related to the diagnosis of eye diseases were researched in books, journals, search engines, print and social media. Resources were cross-checked to verify the information. Artificial intelligence algorithms, some of which were approved by the US Food and Drug Administration, have been adopted in the field of ophthalmology, especially in diagnostic studies. Studies are being conducted that prove that artificial intelligence algorithms can be used in the field of ophthalmology, especially in diabetic retinopathy, age-related macular degeneration, and retinopathy of prematurity. Some of these algorithms have come to the approval stage. The current point in artificial intelligence studies shows that this technology has advanced considerably and shows promise for future work. It is believed that artificial intelligence applications will be effective in identifying patients with preventable vision loss and directing them to physicians, especially in developing countries where there are fewer trained professionals and physicians are difficult to reach. When we consider the possibility that some future artificial intelligence systems may be candidates for moral/ethical status, certain ethical issues arise. Questions about moral/ethical status are important in some areas of applied ethics. Although it is accepted that current intelligence systems do not have moral/ethical status, it has yet to be determined what the exact the characteristics that confer moral/ethical status are or will be.

Keywords:

Artificial intelligence, machine learning, deep learning, ophthalmology, medical ethics

What is Artificial Intelligence?

Artificial intelligence, described simply, is the ability of a computer to mimic the intellectual intelligence unique to humans. This type of intelligence includes qualities such as the ability to link events to specific causes, make generalizations, and learn from experience.¹ As a general public notion, the term is used to describe devices that can provide a reason for a certain phenomenon, develop strategies, make judgments about situations, and have the ability to learn. However, there are ongoing controversies regarding the level and reliability of this intelligence.²

Many different theories on how to evaluate machine intelligence have been proposed. The most famous of these is the Turing test, which was put forward in 1950 by Alan Turing, an English mathematician, computer scientist, and cryptologist. In this test, an assessor compares responses given by a computer and a person without knowing who gave which answer and predicts which one is the machine. If the machine can convince the assessor with its answers at least 30% of the time, it passes the test. In 2014, a program called Eugene Goostman passed this test.³

Yapay Zeka Nedir?

Yapay zeka, basitçe tarif etmek gerekirse, bir bilgisayarın insanlara özgü entelektüel zeka yapısını taklit edebilmesidir. Bu zeka yapısı içinde olayları belirli sebeplerle ilişkilendirebilmek, genelleme yapabilmek, tecrübelerden ders almak gibi özelliklerden bahsedilebilir.¹ Toplumda genel kanı olarak, bir konu hakkında sebep sunabilen, strateji geliştirebilen, durumlar hakkında yargılama yapabilen, öğrenme kabiliyeti olan cihazlarda zeka teriminden bahsedilir. Fakat bu zekanın düzeyi ve güvenilirliği ile ilgili tartışmalar sürmektedir.²

Makinelerin zekasını değerlendirmek üzerine birçok farklı teori sunulmuştur. Bu teorilerden en ünlüsü 1950’de İngiliz matematikçi, bilgisayar bilimcisi ve kriptolog Alan Turing tarafından ortaya atılmış olan “Turing Testi”dir. Bu teste göre değerlendirici pozisyonunda olan kişi, yanıtın kim tarafından verildiğini bilmeden bir bilgisayar ve bir insanın verdiği yanıtları kıyaslar ve hangisinin makine olduğunu tahmin eder. Eğer makine test boyunca verdiği cevaplarla değerlendiriciyi %30 oranında ikna edebilirse testi geçecektir. 2014 yılında Eugene Goostman adında bir program bu testi geçebilmiştir.³

Types of Artificial Intelligence

Artificial intelligence is classified under three headings based on technological achievements and future projections:

1) Artificial Narrow Intelligence: Artificial narrow intelligence, which encompasses nearly all of the software currently described as artificial intelligence, mimics human intelligence within the limited field for which it is designed and responds within this framework.

2) Artificial General Intelligence: This type of artificial intelligence has the same intellectual capacity as humans and is expected in theory to be able to perform tasks at the same level as a person. The consensus among researchers of artificial intelligence is that this type, also called human-level artificial intelligence, must be able to learn and reason, develop strategies, make plans, communicate using language, and synthesize all of these abilities to accomplish a certain task.⁴

3) Artificial Superintelligence: This kind of artificial intelligence is expected to be superior to the most intelligent and talented human brain, and prominent figures in science and technology such as Stephen Hawking and Elon Musk have suggested grim scenarios for the future related to its emergence.

Yapay Zeka Türleri

Yapay zeka, teknolojideki gelişmelere ve yapılan öngörülere göre üç başlık altında sınıflandırılmıştır:

1) Dar yapay zeka (Artificial narrow intelligence): Günümüzde yapay zeka olarak nitelendirilen yazılımların yaklaşık tümünün dahil olduğu dar yapay zeka, kendisi için tanımlanmış sınırlı bir alan içinde, insan zekasını taklit eder ve bu çerçeve içinde yanıt verir.

2) Genel yapay zeka (Artificial general intelligence): İnsanlarla aynı entelektüel kapasitede olup, teorik olarak bir insanın yapabileceği düzeydeki görevleri yerine getirebilmesi beklenen yapay zeka türüdür. İnsan düzeyinde yapay zeka olarak da adlandırılan bu türün; sebep sunmak, strateji geliştirmek, plan yapabilmek, öğrenme kapasitesine sahip olmak, bir dil vasıtasıyla iletişim kurabilmek ve belirli bir görev doğrultusunda tüm bunları sentezleyebilmek kabiliyetlerine sahip olması gerektiği yapay zeka ile ilgilenen araştırmacıların ortak kanısıdır.⁴

3) Süper yapay zeka (Artificial super intelligence): En zeki ve yetenekli insan beyninden daha üstün olacağı beklenmekte olup Stephen Hawking, Elon Musk gibi bilim ve teknolojiyle ilgili önde gelen isimlerin, üzerinde gelecek ile ilgili komplo teorilerini sundukları yapay zeka türüdür.

Artificial Intelligence Learning Algorithms and Ophthalmology

a) Machine Learning: The term “machine learning”, one of the subclasses of artificial intelligence frequently used in ophthalmology studies, was first introduced in 1959 by the engineer Arthur Samuel, a pioneer in artificial intelligence. He defined this term as the ability of machines to learn outcomes that are not explicitly programmed.⁵

In the machine learning technique, the aim is to generate an algorithm based on a certain amount of data entered into a computer and for the computer to then use this algorithm to improve its predictions. The phase in which the device trains with the input to improve its predictions is the learning phase, which is divided into two types: supervised and unsupervised learning. In supervised learning, labels are assigned to the training data as they are entered into the computer, while in unsupervised learning, the device creates its own algorithm from unlabeled input.

b) Deep Learning: As the machine learning technique improves and the amount of input increases, this more advanced method uses multiple layers to generate output, unlike machine learning, which operates with a single layer. Using deep neural networks, the computer can train with much larger data capacity and improve itself with each training cycle to create its own algorith.

Yapay Zekanın Öğrenme Çeşitleri (Algoritmaları) ve Oftalmoloji

a) Makine Öğrenmesi (Machine Learning): Yapay zekanın oftalmoloji alanındaki çalışmalarda sık kullanılan alt sınıflarından biri olan makine öğrenmesi (machine learning) terimi ilk olarak 1959 yılında, yapay zekanın öncü isimlerinden olan mühendis Arthur Samuel tarafından ortaya atılmıştır. Bu terimi “makinelerin bilhassa programlanmadıkları sonuçları öğrenebilme kabiliyeti” olarak açıklamıştır.⁵

Makine öğrenmesi tekniğinde bilgisayara belirli miktarda veri girilerek bu veriler üzerinden bir algoritma oluşturmak ve bu algoritma ile cihazın tahminlerini geliştirmesi hedeflenir. Tahminlerini geliştirmesi için cihaza yüklenen veriler ile cihazın alıştırma yaptığı faz öğrenme fazıdır ve gözetimli öğrenme-gözetimsiz öğrenme olarak iki gruba ayrılır. Gözetimli öğrenmede yüklenen verilere etiketler tanımlanarak bilgisayara tanıtılırken gözetimsiz öğrenmede yüklenen veriler içinden cihaz kendi algoritmasını yaratır.

b) Derin öğrenme (Deep learning): Makine öğrenmesi tekniği geliştikçe ve girilen veriler arttıkça, tek katmanda işlem yaparak verileri tarayan bu teknikten farklı olarak, aynı anda birçok katmanda işlem yaparak bir grup makine öğrenmesi tekniğini aynı anda kullanan, tek seferde sonuca ulaşan daha gelişmiş bir tekniktir. Derin nöral ağlar kullanarak çok daha büyük kapasitede veri ile daha fazla sayıda alıştırma yapabilir ve her alıştırmasında kendini geliştirerek kendi algoritmasını yaratabilir.

Examples of Artificial Intelligence in Medicine

As in many other industries, the use of artificial intelligence in the field of medicine is steadily increasing. Major companies in numerous medical sectors, particularly the pharmaceutical and imaging sectors, have invested billions of dollars in this field, while research on artificial intelligence software is also an area of intense interest in the academic sphere. Although the various publications on the use of artificial intelligence applications in different fields of medicine reveal the breadth of the uses of these techniques, the number of studies that have been approved is still limited.

To give some notable examples of artificial intelligence applications in the field of medicine, in 2016 an artificial intelligence framework by Google called DeepVariant was proven to be able to identify single nucleotide polymorphisms, the most common genetic variation, with 99.9587% accuracy and received an award from the FDA.⁶ The OsteoDetect application, used for wrist fractures in adults, evaluates patients’ X-ray images and was approved by the FDA in 2018.⁷

Artificial intelligence applications developed for purposes such as diagnosing tuberculosis by evaluating a chest X-rays, assessing suspected malignant melanoma based on skin lesion photographs, and detecting lymph node metastasis of breast cancer by analyzing pathology slides, and publications about these represent examples of future areas of use of artificial intelligence.^{8^,}^9,10,11 This is exemplified by a radiology algorithm developed at Stanford University that was able to diagnose pneumonia more accurately than radiologists.¹²

Tıp Alanında Örneklerle Yapay Zeka

Günümüzde pek çok farklı sektörde yaygın olarak kullanılmakta olan yapay zeka uygulamalarının tıp alanında kullanımı da gün geçtikçe artmaktadır. İlaç ve görüntüleme sektörleri başta olmak üzere medikal birçok sektördeki dev firmaların bu konudaki yatırımları milyar dolarlarla ifade edilirken, akademik anlamda da yapay zeka yazılımları ile ilgili çalışmalar yoğun ilgi çekmektedir. Yapay zeka uygulamalarının tıbbın farklı alanlarındaki kullanımlarına dair çeşitli yayınlar bu uygulamaların kullanım alanlarının genişliğini gözler önüne serse de onay alan çalışmalar henüz sınırlı sayıdadır.

Tıp alanındaki yapay zeka uygulamalarına verilebilecek dikkat çekici örneklerden bazılarına değinecek olursak; 2016’da Google’ın DeepVariant adlı yapay zeka uygulamasının, en sık genetik varyasyon olan tek nükleotid polimorfizmini %99,9587 doğrulukta belirleyebildiği kanıtlanmış ve FDA’dan ödül almıştır.⁶

Yetişkinlerde el bileği kırıklarında kullanılmak üzere hastaların X-ray görüntülerini değerlendirerek çalışan OsteoDetect adlı uygulama 2018 yılında FDA onayı almıştır.⁷

Akciğer grafisi değerlendirerek tüberküloz tanısı koymak, deri lezyonu fotoğraflarından malign melanom şüphesi değerlendirmek, patoloji preparatlarına bakarak meme kanseri lenf nodu metastazı varlığını tespit etmekgibi konularda geliştirilen yapay zeka uygulamaları ve bunlarla ilgili yayınlar yapay zekanın gelecekteki kullanım alanlarına örnek teşkil etmektedir.^8,9,10,11 Öyle ki, Stanford Üniversitesi’nde geliştirilen radyoloji algoritması, pnömoni tanısında radyologlardan daha kesin tanılar koyabilmiştir.¹²

What Can Artificial Intelligence Do for Ophthalmology?

The field of ophthalmology is well suited for artificial intelligence studies, with its numerous digital techniques such as color fundus photography, optical coherence tomography (OCT), and computerized visual field (VF) testing and the huge databases they provide.

In addition to this, the global increase in life expectancy is accompanied by an increase in eye diseases that cause preventable vision loss.^{13^,}¹⁴ Solutions are sought for the early diagnosis and treatment of these diseases, especially in regions where access to physicians is difficult. Artificial intelligence applications are being developed for many different eye diseases, particularly diabetic retinopathy (DR), age-related macular degeneration (AMD), glaucoma, and retinopathy of prematurity (ROP), which are the leading causes of vision loss.¹⁵

Oftalmolojide Yapay Zeka Ne Yapabilir?

Oftalmoloji; renkli fundus fotoğrafları, optik koherens tomografi (OKT), bilgisayarlı görme alanı (visual field-VF) gibi farklı dijital teknikler ile yapay zeka çalışmaları için uygun ve geniş veritabanına sahip bir alandır. Buna ek olarak dünyada artan ortalama yaşam süresi, önlenebilir görme kaybına sebep olan göz hastalıklarının artışını beraberinde getirmektedir.^13,14

Bu hastalıkların erken teşhisi ve tedavisi aşamasında, bilhassa hekime ulaşımın zor olduğu bölgelerde çözüm yolları aranmaktadır. Özellikle toplumda en yaygın görme kaybı sebepleri olan diyabetik retinopati (DR), yaşa bağlı maküla dejenerasyonu (YBMD), glokom ve prematüre retinopatisi (ROP) başlıkları ön planda olmak üzere birçok farklı göz hastalığı için yapay zeka uygulamaları oluşturulmaktadır.¹⁵

Artificial Intelligence and Diabetic Retinopathy

Due to the rapidly increasing number of patients worldwide, DR has generated the most interest in terms of the use of artificial intelligence in ophthalmology. IDx-DR, the first FDA-approved device using artificial intelligence software, was also developed for this area.¹⁶

The IDx-DR uses a Topcon NW400 fundus camera to classify patients according to retinopathy level. Ease of use was cited as the priority when choosing the fundus camera. The operators selected to obtain the fundus photographs had no previous experience using a fundus camera. Patients were grouped into those with mild to advanced DR according to the American Academy of Ophthalmology classification (Preferred Practice Patterns for Diabetic Retinopathy) and those without retinopathy, and the patients were recommended follow-up examination at 12 months or sooner according to their results. A total of 900 patients participated in the study and the sensitivity and specificity of the device were found to be 87.4% and 89.5%, respectively. The device began to be used at the University of Iowa in 2018 after receiving FDA approval.¹⁶

The IDx-DR was developed using software that uses deep learning techniques, and there are a growing number of similar studies using fundus cameras and deep learning software.^{17^,}^18,19,20 Thanks to deep learning applications, it is possible to develop software with databases containing over 100,000 data points.^{21^,}²²

There are examples of studies using machine learning methods with fundus photographs, machine learning with OCT, and deep learning methods with OCT.^{19^,}^{23,24,25,26,27} Some of these studies have reported nearly 100% sensitivity or specificity rates.^{28^,}²⁹

Yapay Zeka ve Diyabetik Retinopati

Diyabetik retinopati, dünya çapında hızla artan hasta sayısı nedeniyleoftalmoloji alanında yapay zekanın kullanımıyla ilgili en çok ilgi gören alandır. Yapay zeka yazılımı kullanan ilk FDA onaylı cihaz olan IDx-DR’de bu alanda geliştirilmiştir.¹⁶

IDx-DR, Topcon NW400 fundus kamerası kullanarak, hastaları retinopati seviyesine göre sınıflandırmaktadır. Fundus kamerası seçiminde ise kullanım kolaylığının ön plana alındığı belirtilmektedir. Fundus fotoğraflarını çeken operatörler daha önce fundus kamerası kullanmamış kişiler arasından seçilmiştir. Hastalar, Amerikan Oftalmoloji Akademisi Sınıflaması’na (Preferred Practice Patterns for Diabetic Retinopathy) göre hafif diyabetik retinopatiden ileri retinopatisi olanlar ve olmayanlar olarak ikiye ayrılarak, sonuçlara göre hastalara 12 ay sonra tekrar kontrol ya da yakın zamanda oftalmolojik muayene önerilmiştir. Çalışmaya 900 hasta katılmış ve çalışma sonucunda cihazın duyarlılığı %87,4 ve özgüllüğü %89,5 olarak bulunmuştur. FDA onayı sonrası cihaz 2018 yılında Iowa Üniversitesi’nde kullanılmaya başlanmıştır.¹⁶

IDx-DR, derin öğrenme uygulamaları kullanan bir yazılım kullanılarak geliştirilmiştir, buna benzer olarak fundus kamerası ve derin öğrenme yazılımı kullanan farklı çalışma örnekleri artmaktadır.^17,18,19,20 Derin öğrenme uygulamaları sayesinde veri tabanı yüz bini aşan yazılımlar geliştirilebilmektedir.^21,22

Fundus fotoğrafı üzerinden çalışılarak makine öğrenmesi uygulamaları kullanılan, OCT ile çalışılarak makine öğrenmesi kullanılanve OCT ile çalışılarak derin öğrenme uygulamaları kullanılan çalışma örnekleri bulunmaktadır.^{19,23,24,25,26,27} Bu çalışmalarda %100’e yakın duyarlılık ya da özgüllük oranları bildirilenler bulunmaktadır.^28,29

Artificial Intelligence and Age-related Macular Degeneration

As with DR, an increasing number of studies are investigating software that uses artificial intelligence for the early diagnosis and classification of AMD. The earliest published studies involved software developed using fundus photography and machine learning with database sizes smaller than 1,000.^{30^,}^31,32 Later, with software using deep learning technology, database sizes increased and high sensitivity and specificity rates were reached.^{20^,}^33,34

Ting et al.²⁰ used a database of 72,610 fundus photographs and classified patients as those with intermediate to advanced AMD and those without according to the AREDS (Age-Related Eye Disease Study) classification. They reported sensitivity and specificity of 93.2% and 88.2%, respectively.²⁰

Burlina et al.³³ classified patients with software developed using 130,000 images from 4613 patients and reporting a 91.6% accuracy rate in identifying those with moderate and advanced AMD patients.

Grassmann et al.³⁴ tested an algorithm they generated from 120,656 fundus photographs of 3,654 patients against the AREDS database and reported an accuracy rate of 84.2% in differentiating early and late disease and 94.3% accuracy in identifying healthy subjects.

Yapay Zeka ve Yaşa Bağlı Maküla Dejenerasyonu

Diabetik retinopatide olduğu gibi, YBMD’de da erken tanı ve sınıflandırma amacıyla yapay zeka kullanılarak geliştirilen yazılımlar ile ilgili çalışmalar artmaktadır. Önceleri fundus fotoğrafı ve makine öğrenmesi kullanarak geliştirilen, veri tabanı boyutları 1000’den az olan çalışmalar yayınlanmıştır.^30,31,32 Daha sonra derin öğrenme teknolojisi kullanan yazılımlar ile veri tabanı boyutları artmış, yüksek duyarlılık ve özgüllük oranlarına ulaşılmıştır.^20,33,34

Ting ve ark.²⁰ 72.610 fundus fotoğrafını veri tabanı olarak kullanmış ve hastaları AREDS (Age-Related Eye Disease Study) Sınıflaması’na göre orta düzey ve üstü olanlar ve olmayanlar olarak sınıflandırmıştır. Çalışma sonucunda duyarlılık %93,2, özgüllük %88,2 olarak belirtilmiştir.

Burlina ve ark.³³ 4.613 hastanın 130.000 görüntüsünü kullanarak oluşturdukları yazılım ile hastaları sınıflandırmış, orta ve ileri düzey YBMD hastalarını belirlemede %91,6 doğruluk oranı belirtmişlerdir.

Grassmann ve ark.’da³⁴ 3.654 hastanın 120.656 fundus fotoğrafını tarayarak oluşturdukları algoritmalarını AREDS veri tabanına karşı test etmişler ve hastalığı erken ve geç dönem olarak sınıflandırmada %84,2 doğruluk oranı; sağlıklı hastaları belirlemekte ise %94,3 doğruluk oranı belirtmişlerdir.

Artificial Intelligence and Glaucoma

Glaucoma is among the leading causes of vision loss worldwide and has also attracted the attention of artificial intelligence researchers due to the importance of its early diagnosis and treatment.¹⁵

Initially, studies using machine learning to identify glaucomatous optic nerve damage based on fundus photographs were published.^{35^,}^36,37 These were followed by studies that used deep learning technology with much larger databases compared to the earlier machine learning studies.^{20^,}^38,39 In another study using a database of 125,189 fundus photographs, Ting et al.²⁰ reported a sensitivity of 96.4% and specificity of 87.2%.

Studies are being conducted on the use of imaging modalities other than fundus photography in the diagnosis and monitoring of glaucoma. In addition to artificial intelligence applications created using computerized VF and OCT data, studies have also been published describing programs that are able to evaluate patients based on data from both of these examination devices.^{40^,}^{41,42,43,44,45}

Yapay Zeka ve Glokom

Glokom, dünya çapında en yaygın görme kaybı sebeplerinden olup, erken tanı ve tedavisinin önemi nedeniyle yapay zeka araştırmacılarının dikkatini yoğun olarak çeken bir diğer konudur.¹⁵

Fundus fotoğrafları üzerinden glokomatöz optik nöron hasarını belirlemek üzere makine öğrenmesi yöntemini kullanan çalışmalar yayınlanmıştır.^35,36,37 Bu çalışmalardan sonra derin öğrenme teknolojisi kullanan ve veri tabanı boyutları makine öğrenmesi kullanılan çalışmalara göre oldukça büyük olan çalışmalar üzerine yoğunlaşılmıştır.^20,38,39 Ting ve ark.’nın²⁰ veri tabanında 125.189 fundus fotoğrafını kullandıkları çalışmada duyarlılık %96,4, özgüllük %87,2 olarak belirtilmiştir.

Fundus fotoğrafları dışında glokom tanı ve takibinde kullanılan diğer görüntüleme tetkikleri ile ilgili çalışmalar yapılmaktadır. Bilgisayarlı görme alanı verileri ve OCT verilerikullanılarak hazırlanan yapay zeka uygulamalarının yanında, bu iki tetkik cihazından da veri alınarak değerlendirme yapabilen uygulama çalışmaları yayınlanmıştır.^{40,41,42,43,44,45}

Artificial Intelligence and Retinopathy of Prematurity

ROP is a leading cause of vision loss in childhood worldwide and its prevalence is reported as 6-18% in different studies.⁴⁶ According to the ETROP (Early Treatment for Retinopathy of Prematurity) study, early treatment is vital for improving visual acuity, and 9% of patients have permanent vision loss despite early treatment.⁴⁷

Although the impact of ROP diagnosis and treatment on patients’ visual acuity outcomes and quality of life is known, access to physicians specializing in ROP can be limited, especially in less developed countries. One of the reasons for this is that the follow-up and treatment of ROP requires a long time and specialized education, even for ophthalmologists. This coupled with high malpractice rates and lawsuits result in physicians avoiding this area.^{48^,}⁴⁹ In addition, parameters used in the diagnosis of ROP, such as zone, stage, and presence of additional diseases lead to diagnostic variations even among ROP specialists.⁴⁸ The difficulty in finding specialists and the diagnostic variations among specialists has prompted artificial intelligence researchers to conduct studies on ROP.

Brown et al.⁵⁰ developed software using deep learning technology and a database of 5,511 fundus images obtained with a RetCam fundus camera and reported 93% sensitivity and 94% specificity in determining the presence of additional disease.

In an application developed by Redd et al.⁵¹ based on the same deep learning technology, the software was found to have 0.96 and 0.91 area under the curve values, respectively, in the identification of type 1 ROP and clinically significant ROP.

Yapay Zeka ve Prematüre Retinopatisi

Prematüre retinopatisi, tüm dünyada çocukluk çağı görme kaybı nedenleri arasında önde gelmektedir ve farklı çalışmalarda %6-18 arasında oranlar belirtilmiştir.⁴⁶ ETROP (Early Treatment for Retinopathy of Prematurity) çalışmasına göre erken tedavi görme keskinliğini arttırmakta önemlidir ve erken tedaviye rağmen, hastaların %9’unda kalıcı görme kaybı gelişmektedir.⁴⁷

ROP tanı ve tedavisinin hastanın sonuç görme keskinliği ve yaşam kalitesi üzerine etkisinin önemi bilinmesine rağmen, özellikle gelişmemiş ülkelerde ROP ile ilgilenen hekimlere ulaşmak zordur. Bunun bir sebebi de ROP takip ve tedavisinde, oftalmologlar için dahi uzun bir zaman ve eğitim süreci gerekliliği olması, bunun yanında yüksek malpraktis oranları ve dava açılması nedeniyle hekimlerin bu alandan uzaklaşmasıdır.^48,49 Bunlara ek olarak, ROP tanısında kullanılan zon, evre ve artı hastalık varlığı gibi parametreler, ROP ile ilgilenen uzmanlar arasında bile tanısal varyasyonlara yol açmaktadır.⁴⁸ Uzman hekime ulaşmaktaki güçlükler ve uzmanlar arası tanısal varyasyonlar, yapay zeka araştırmacılarını ROP konusunda çalışmaya yöneltmektedir.

Brown ve ark.’nın⁵⁰ RetCam fundus kamerasıyla çekilmiş 5.511 fundus görüntüsünü veri tabanı olarak kullandıkları derin öğrenme teknolojisi kullanan yazılımlarının, artı hastalık varlığını belirlemekte %93 duyarlılık ve %94 özgüllüğe sahip olarak belirtilmiştir.

Redd ve ark.’nın⁵¹ aynı derin öğrenme teknolojisi üzerine geliştirdikleri uygulamada ise yazılımın tip 1 ROP ve klinik olarak anlamlı ROP’yi belirlemekte sırasıyla 0,96 ve 0,91 eğri altı alan değerlerine sahip olduğu bildirilmiştir.

Other Applications in Ophthalmology

Data from a study by De Fauw et al.⁵² conducted at Moorfields Eye Hospital using Google’s deep learning technology called DeepMind presents a striking illustration of the level artificial intelligence has reached.

The artificial intelligence algorithm was trained by introducing 10 different lesions such as hemorrhage and fluid and using 14,884 untagged OCT images, and is able to distinguish more than 50 retinal diseases. The study included data obtained from 37 different OCT devices from different centers affiliated with Moorfields, and data from the application were compared with the decisions of four ophthalmologists and four optometrists affiliated with Moorfields. These physicians and the device were asked to classify OCT images based on the need for referral as urgent, semi-urgent, routine examination, and observation.

At the end of the study, the software’s error rate (5.5%) was comparable to those of the hospital’s two best retina specialists (6.7% and 6.8%) and was significantly better than those the other six specialists (10-24.1%). In particular, it was reported that the software made no errors in the urgent referral group.⁵²

Oftalmolojide Diğer Uygulamalar

De Fouw ark.’nın⁵² Google’ın DeepMind adlı derin öğrenme teknolojisini kullanarak Moorfields Göz Hastanesi’nde yaptıkları çalışmanın verileri, yapay zekanın gelmiş olduğu seviyeyi çarpıcı bir şekilde göstermektedir.

Yapay zeka algoritması; kanama, sıvı gibi 10 farklı lezyon tanıtılarak 14.884 etiketlenmemiş OCT görüntüsünü kullanılarak eğitilmiştir ve 50’den fazla retina hastalığını ayırt edebilmektedir. Çalışmada Moorfields’a bağlı farklı merkezlerden 37 farklı OCT cihazından alınan veriler kullanılmış ve uygulamanın verileri, Moorfields’a bağlı dört oftalmolog ve dört optometristin kararları ile karşılaştırılmıştır. Cihazdan ve kişilerden OCT görüntülerini muayene ihtiyacına göre acil, kısmen acil, rutin muayene ve gözlem olarak sınıflandırmaları istenmiştir.

Çalışma sonucunda, cihazın hata oranı (%5,5), hastanenin en iyi iki retina uzmanı ile kıyaslanabilir düzeyde (%6,7 ve %6,8) ve diğer altı uzmandan anlamlı olarak daha iyi (%10 - %24,1) olarak belirtilmiştir. Özellikle acil olarak sınıflandırılan grup içinde cihazın hata yaptığı olgu olmadığı belirtilmektedir.⁵²

Discussion

The progress made in artificial intelligence studies shows that important advances in this technology are ongoing, and it is clear that potential future applications lie on the horizon and are promising for future studies. It is believed that artificial intelligence will be effective in identifying patients with preventable vision loss and referring them to a physician, especially in developing countries where access to physicians is difficult and the population of trained individuals is low.

The diagnostic spectrum in which artificial intelligence may be used and its possible clinical benefits represent a broad field of study. With technologies similar to those in previous studies, applications based on different imaging modalities can be developed in various areas such as occlusive vascular disease, keratoconus, and retinitis pigmentosa. Beyond just screening for or diagnosing diseases, surgical applications may be created to provide guidance for physicians in areas such as determining the ideal type of intraocular lens for a patient or estimating the risk of surgery.

Another important aspect of artificial intelligence that should be discussed is the potential ethical conflicts. Imagine that in the near future, a bank evaluates home loan applications using machine learning algorithms. Now imagine that someone denied a loan sues the bank, claiming that the algorithm racially discriminated in its evaluation. The bank will state that this is impossible because the algorithm does not know the applicants’ race. In fact, the reason the bank implemented such a system was to eliminate unpleasant situations such as the involvement of human emotions. Nevertheless, suppose statistics indicate that the bank’s approval rate is steadily falling for black applicants. When ten equivalent applications are entered into the system, the algorithm accepts those of the white applicants and rejects those of the black applicants. What do you think is happening here? It may not be easy to determine the answer. If the machine learning algorithm is created using a complex artificial neural network or based on a genetic algorithm generated by controlled evolution, it would be impossible to understand why and according to what data the algorithm makes its racially discriminant decisions. In contrast, decision trees or networks in special computer language are much more transparent in terms of allowing the programmer to analyze them. This may allow an auditor to discover, for example, that the decision is reached based on where the applicants were born or the fact that they previously resided mostly in suburban neighborhoods, i.e., their addresses.

Artificial intelligence algorithms play an increasingly prominent role in modern society. In general, however, those who are affected by them are not even aware that such a thing exists in the background.

When we consider the possibility that some future artificial intelligence systems may be candidates for moral/ethical status, various ethical issues arise. Relations with beings of moral/ethical status are not entirely a matter of rationality; we also have moral/ethical reasons to treat them in certain ways and to avoid mistreating them. Kamm⁵³ proposed the following definition of moral status that will serve our purpose: An entity has moral status when it is morally important in its own right and some things are morally permissible or impermissible to do to them, for their own sake.

Questions about moral/ethical status are important in some areas of applied ethics. For example, disputes regarding the moral acceptability of abortion generally influence disagreements about the moral/ethical status of the human embryo. The controversies related to animal experimentation and the treatment of animals in the food industry include questions about the moral/ethical status of different animal species. Our obligations to persons with severe dementia, such as end-stage Alzheimer’s patients, may also depend on questions of moral/ethical status.

Current artificial intelligence systems are generally regarded as not having moral status. At least as far as the programs themselves are concerned, we can modify, copy, terminate, delete, and use computer programs as desired. The moral/ethical restrictions involved in our relationships with contemporary artificial intelligence systems are based on our obligations to other beings, such as human race itself. However, we have no duty to the systems themselves.

Although there seems to be a consensus that current artificial intelligence systems do not have moral/ethical status, it is not clear what the characteristics determining moral/ethical status are or will be. In addition, infants and individuals suffering from severe mental illnesses do not meet the criteria for cognitive capacity. Some authorities do not regard people with mental illness as having full moral status.

Conclusion

The discipline of artificial intelligence ethics, especially considering artificial general intelligence, differs fundamentally from the moral/ethical discipline of non-cognitive technologies:

• The local, specific behavior of artificial intelligence may not be predictable apart from its safety, even if programmers do everything right.

• Verifying the reliability of the system can become a greater challenge, as it requires verifying what the system is trying to do rather than verifying the safe behavior of the system in all areas in which it operates.

• Ethical cognition should be addressed as an engineering issue.^{59^,}⁶⁰

Ancient civilizations considered slavery acceptable; we believe otherwise. Ethical debates over voting rights for women and blacks continued even into the nineteenth and twentieth centuries. Advancing science and increasing technological capabilities are not the only differences between modern and ancient civilizations. There is also a difference in ethical perspective. It is very likely that machine ethics will present our greatest challenge. The question then becomes:

How will you create artificial intelligence that, as it operates, will eventually become more ethical than you?

References

Copeland BJ. Artificial intelligence | Definition, Examples, and Applications | Britannica.com. Encyclopedia Britannica, https://www.britannica.com/technology/artificial-intelligence (2019, accessed 25 February 2019).

Adams S, Arel I, Bach J, Coop R, Furlan R, Goertzel B, Hall JS, Samsonovich A, Scheutz M, Schlesinger M, Shapiro SC, Sowa J. Mapping the Landscape of Human-Level Artificial General Intelligence. AI Magazine. 2012;33:25-42.

Computer AI passes Turing test in ‘world first’ - BBC News, https://www.bbc.com/news/technology-27762088 (2014, accessed 25 February 2019).

Russell SJ, Norvig P. Artificial intelligence : a modern approach. 3rd ed. New Jersey: Pearson Education. 2010.

Samuel AL. Some Studies in Machine Learning Using the Game of Checkers. II—Recent Progress. In: Computer Games I. New York, NY: Springer New York, pp. 366-400.

PrecisionFDA Truth Challenge - Google Genomics v1 documentation, https://googlegenomics.readthedocs.io/en/staging-2/use_cases/discover_public_data/precision_fda.html (accessed 26 February 2019).

Levy MC. FDA and Artificial Intelligence in Digital Health Innovation | Artificial Intelligence Law Blog, https://www.artificialintelligencelawblog.com/2018/12/fda-artificial-intelligence-digital-health-innovation/ (2018, accessed 3 March 2019).

Lakhani P, Sundaram B. Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks. Radiology. 2017;284:574-582.

Ting DSW, Yi PH, Hui F. Clinical Applicability of Deep Learning System in Detecting Tuberculosis with Chest Radiography. Radiology. 2018;286:729-731.

Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017;542:115-118.

Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, van der Laak JAWM; the CAMELYON16 Consortium, Hermsen M, Manson QF, Balkenhol M, Geessink O, Stathonikos N, van Dijk MC, Bult P, Beca F, Beck AH, Wang D, Khosla A, Gargeya R, Irshad H, Zhong A, Dou Q , Li Q, Chen H, Lin HJ, Heng PA, Haß C, Bruni E, Wong Q, Halici U, Öner MÜ, Cetin-Atalay R, Berseth M, Khvatkov V, Vylegzhanin A, Kraus O, Shaban M, Rajpoot N, Awan R, Sirinukunwattana K, Qaiser T, Tsang YW, Tellez D, Annuscheit J, Hufnagl P, Valkonen M, Kartasalo K, Latonen L, Ruusuvuori P, Liimatainen K, Albarqouni S, Mungal B, George A, Demirci S, Navab N, Watanabe S, Seno S, Takenaka Y, Matsuda H, Ahmady Phoulady H, Kovalev V, Kalinovsky A, Liauchuk V, Bueno G, Fernandez-Carrobles MM, Serrano I, Deniz O, Racoceanu D, Venâncio R.Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA. 2017;318:2199-2210.

Kubota T. Algorithm better at diagnosing pneumonia than radiologists | News Center | Stanford Medicine, https://med.stanford.edu/news/all-news/2017/11/algorithm-can-diagnose-pneumonia-better-than-radiologists.html (accessed 3 March 2019).

Arias E, Heron M, Xu J. United States life tables, 2014. Natl Vital Stat Rep. 2017;66:1-64.

Zheng Y, He M, Congdon N. The worldwide epidemic of diabetic retinopathy. Indian J Ophthalmol. 2012;60:428-431.

Flaxman SR, Bourne RRA, Resnikoff S, Ackland P, Braithwaite T, Cicinelli MV, Das A, Jonas JB, Keeffe J, Kempen JH, Leasher J, Limburg H, Naidoo K, Pesudovs K, Silvester A, Stevens GA, Tahhan N, Wong TY, Taylor HR; Vision Loss Expert Group of the Global Burden of Disease Study. Global causes of blindness and distance vision impairment 1990-2020: a systematic review and meta-analysis. Lancet Glob Health. 2017;5:e1221-e1234.

Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digital Medicine. 2018;1:39.

Ramachandran N, Hong SC, Sime MJ, Wilson AG. Diabetic retinopathy screening using deep neural network. Clinical & Experimental Ophthalmology. 2018;46:412-416.

Gargeya R, Leng T. Automated Identification of Diabetic Retinopathy Using Deep Learning. Ophthalmology. 2017;124:962-969.

Abràmoff MD, Lou Y, Erginay A, Clarida W, Amelon R, Folk JC, Niemeijer M. Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning. Invest Ophthalmol Vis Sci. 2016;57:5200-5206.

Ting DSW, Cheung CY, Lim G, Tan GSW, Quang ND, Gan A, Hamzah H, Garcia-Franco R, San Yeo IY, Lee SY, Wong EYM, Sabanayagam C, Baskaran M, Ibrahim F, Tan NC, Finkelstein EA, Lamoureux EL, Wong IY, Bressler NM, Sivaprasad S, Varma R, Jonas JB, He MG, Cheng CY, Cheung GCM, Aung T, Hsu W, Lee ML, Wong TY. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. JAMA. 2017;318:2211-2223.

Gulshan V, Peng L, Coram M. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316:2402-2410.

Li Z, Keel S, Liu C, He Y, Meng W, Scheetz J, Lee PY, Shaw J, Ting D, Wong TY, Taylor H, Chang R, He M. An Automated Grading System for Detection of Vision-Threatening Referable Diabetic Retinopathy on the Basis of Color Fundus Photographs. Diabetes Care. 2018;41:2509-2516.

Acharya UR, Ng EYK, Tan J-H, Sree SV, Ng KH. An integrated index for the identification of diabetic retinopathy stages using texture parameters. J Med Syst. 2012;36:2011-2020.

Ganesan K, Martis RJ, Acharya UR, Chua CK, Min LC, Ng EY, Laude A. Computer-aided diabetic retinopathy detection using trace transforms on digital fundus images. Med Biol Eng Comput. 2014;52:663-672.

Sandhu HS, Eltanboly A, Shalaby A, Keynton RS, Schaal S, El-Baz A. Automated Diagnosis and Grading of Diabetic Retinopathy Using Optical Coherence Tomography. Invest Ophthalmol Vis Sci. 2018;59:3155-3160.

Adhi M, Semy SK, Stein DW, Potter DM, Kuklinski WS, Sleeper HA, Duker JS, Waheed NK. Application of Novel Software Algorithms to Spectral-Domain Optical Coherence Tomography for Automated Detection of Diabetic Retinopathy. Ophthalmic Surg Lasers Imaging Retina. 2016;47:410-417.

ElTanboly A, Ismail M, Shalaby A, Switala A, El-Baz A, Schaal S, Gimel’farb G, El-Azab M. A computer-aided diagnostic system for detecting diabetic retinopathy in optical coherence tomography images. Med Phys. 2017;44:914-923.

Qi SR. Machine Learning and OCT Images — the Future of Ophthalmology, https://medium.com/health-ai/machine-learning-and-oct-images-the-future-of-ophthalmology-47dc64ee9dc6.

Qi SR. Deep Learning in Ophthalmology — How Google Did It – Health, https://medium.com/health-ai/deep-learning-in-ophthalmology-using-128-175-retinal-images-59814e8a3f68 (accessed 4 March 2019).

Agurto C, Barriga ES, Murray V, Nemeth S, Crammer R, Bauman W, Zamora G, Pattichis MS, Soliz P. Automatic Detection of Diabetic Retinopathy and Age-Related Macular Degeneration in Digital Fundus Images. Invest Ophthalmol Vis Sci. 2011;52:5862-5871.

Zheng Y, Hijazi MHA, Coenen F. Automated “Disease/No Disease” Grading of Age-Related Macular Degeneration by an Image Mining Approach. Invest Ophthalmol Vis Sci. 2012;53:8310-8318.

Mookiah MR, Acharya UR, Koh JE, Chandran V, Chua CK, Tan JH, Lim CM, Ng EY, Noronha K, Tong L, Laude A. Automated diagnosis of Age-related Macular Degeneration using greyscale features from digital fundus images. Comput Biol Med. 2014;53:55-64.

Burlina PM, Joshi N, Pekala M, Pacheco KD, Freund DE, Bressler NM. Automated Grading of Age-Related Macular Degeneration From Color Fundus Images Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2017;135:1170-1176.

Grassmann F, Mengelkamp J, Brandl C, Harsch S, Zimmermann ME, Linkohr B, Peters A, Heid IM, Palm C, Weber BHF. A Deep Learning Algorithm for Prediction of Age-Related Eye Disease Study Severity Scale for Age-Related Macular Degeneration from Color Fundus Photography. Ophthalmology. 2018;125:1410-1420.

Singh A, Dutta MK, ParthaSarathi M, Uher V, Burget R. Image processing based automatic diagnosis of glaucoma using wavelet features of segmented optic disc from fundus image. Comput Methods Programs Biomed. 2016;124:108-120.

Salam AA, Khalil T, Akram MU, Jameel A, Basit I. Automated detection of glaucoma using structural and non structural features. Springerplus. 2016;5:1519.

Chakrabarty L, Joshi GD, Chakravarty A, Raman GV, Krishnadas SR, Sivaswamy J. Automated Detection of Glaucoma From Topographic Features of the Optic Nerve Head in Color Fundus Photographs. J Glaucoma. 2016;25:590-597.

Chen X, Xu Y, Kee Wong DW, Wong TY, Liu J. Glaucoma detection based on deep convolutional neural network. In: 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Conference 2015:715-718.

Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. Ophthalmology. 2018;125:1199-1206.

Muhammad H, Fuchs TJ, De Cuir N, De Moraes CG, Blumberg DM, Liebmann JM, Ritch R, Hood DC. Hybrid Deep Learning on Single Wide-field Optical Coherence tomography Scans Accurately Classifies Glaucoma Suspects. J Glaucoma. 2017;26:1086-1094.

Oh E, Yoo TK, Hong S. Artificial Neural Network Approach for Differentiating Open-Angle Glaucoma From Glaucoma Suspect Without a Visual Field Test. Invest Ophthalmol Vis Sci. 2015;56:3957-3966.

Bowd C, Lee I, Goldbaum MH, Balasubramanian M, Medeiros FA, Zangwill LM, Girkin CA, Liebmann JM, Weinreb RN. Predicting Glaucomatous Progression in Glaucoma Suspect Eyes Using Relevance Vector Machine Classifiers for Combined Structural and Functional Measurements. Invest Ophthalmol Vis Sci. 2012;53:2382-2389.

Niwas SI, Lin W, Bai X, Kwoh CK, Jay Kuo CC, Sng CC, Aquino MC, Chew PT. Automated anterior segment OCT image analysis for Angle Closure Glaucoma mechanisms classification. Comput Methods Programs Biomed. 2016;130:65-75.

Silva FR, Vidotti VG, Cremasco F, Dias M, Gomi ES, Costa VP. Sensitivity and specificity of machine learning classifiers for glaucoma diagnosis using Spectral Domain OCT and standard automated perimetry. Arq Bras Oftalmol. 2013;76:170-174.

Vidotti VG, Costa VP, Silva FR, Resende GM, Cremasco F, Dias M, Gomi ES. Sensitivity and Specificity of Machine Learning Classifiers and Spectral Domain OCT for the Diagnosis of Glaucoma. Eur J Ophthalmol. 2012;23:61-69.

Fleck BW, Dangata Y. Causes of visual handicap in the Royal Blind School, Edinburgh, 1991-2. Br J Ophthalmol. 1994;78:421.

Early Treatment for Retinopathy of Prematurity Cooperative Group, Good WV, Hardy RJ, Dobson V, Palmer EA, Phelps DL, Tung B, Redford M. Final visual acuity results in the early treatment for retinopathy of prematurity study. Arch Ophthalmol. 2010;128:663-671.

Chiang MF, Jiang L, Gelman R, Du YE, Flynn JT. Interexpert Agreement of Plus Disease Diagnosis in Retinopathy of Prematurity. Arch Ophthalmol. 2007;125:875-880.

Braverman RS, Enzenauer RW. Socioeconomics of Retinopathy of Prematurity In-Hospital Care. Arch Ophthalmol. 2010;128:1055-1058.

Brown JM, Campbell JP, Beers A, Chang K, Ostmo S, Chan RVP, Dy J, Erdogmus D, Ioannidis S, Kalpathy-Cramer J, Chiang MF, Imaging and Informatics in Retinopathy of Prematurity (i-ROP) Research Consortium. Automated Diagnosis of Plus Disease in Retinopathy of Prematurity Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2018;136:803-810.

Redd TK, Campbell JP, Brown JM, Kim SJ, Ostmo S, Chan RVP, Dy J, Erdogmus D, Ioannidis S, Kalpathy-Cramer J, Chiang MF; Imaging and Informatics in Retinopathy of Prematurity (i-ROP) Research Consortium. Evaluation of a deep learning image assessment system for detecting severe retinopathy of prematurity. Br J Ophthalmol. 2018;bjophthalmol-2018-313156.

De Fauw J, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N, Blackwell S, Askham H, Glorot X, O’Donoghue B, Visentin D, van den Driessche G, Lakshminarayanan B, Meyer C, Mackinder F, Bouton S, Ayoub K, Chopra R, King D, Karthikesalingam A, Hughes CO, Raine R, Hughes J, Sim DA, Egan C, Tufail A, Montgomery H, Hassabis D, Rees G, Back T, Khaw PT, Suleyman M, Cornebise J, Keane PA, Ronneberger O. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med. 2018;24:1342-1350.

Kamm FM. Terrorism and several moral distinctions. Legal Theory. 2006;12:19-69.

Frankish K, Ramsey WM. The Cambridge handbook of artificial intelligence. Choice Reviews Online. 2015;52:3019-3019.

Keskinbora HK, Keskinbora K. Ethical considerations on novel neuronal interfaces. Neurol Sci. 2018;39:607-613.

Bostrom N, Cirkovic MM. Global catastrophic risks. Choice Reviews Online. 2013;46:6152-6152.

Vinge V. The Coming Technological Singularity: How to Survive in the Post-Human Era. Science Fiction Criticism: An Anthology of Essential Writings. 1993;352-363.

Kurzweil R. The singularity is near : when humans transcend biology. Viking; 2005. https://www.amazon.com/Singularity-Near-Humans-Transcend-Biology/dp/0143037889

Frankish K, Ramsey WM. The Cambridge Handbook of Artificial Intelligence, https://philpapers.org/rec/FRATCH-2 (2014, accessed 4 March 2019).

Keskinbora KH, Jameel MA. Nanotechnology Applications and Approaches in Medicine: A Review. Journal of Nanotechnology; 2018:1-5.

This page is for health professionals only.

Artificial Intelligence and Ophthalmology

ABSTRACT

Discussion

Conclusion

References