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

Abstract

<jats:title>Abstract</jats:title> <jats:sec> <jats:title>Background</jats:title> <jats:p>Glaucoma is the most frequent cause of irreversible blindness worldwide. There is no cure, but early detection and treatment can slow the progression and prevent loss of vision. It has been suggested that artificial intelligence (AI) has potential application for detection and management of glaucoma.</jats:p> </jats:sec> <jats:sec> <jats:title>Sources of data</jats:title> <jats:p>This literature review is based on articles published in peer-reviewed journals.</jats:p> </jats:sec> <jats:sec> <jats:title>Areas of agreement</jats:title> <jats:p>There have been significant advances in both AI and imaging techniques that are able to identify the early signs of glaucomatous damage. Machine and deep learning algorithms show capabilities equivalent to human experts, if not superior.</jats:p> </jats:sec> <jats:sec> <jats:title>Areas of controversy</jats:title> <jats:p>Concerns that the increased reliance on AI may lead to deskilling of clinicians.</jats:p> </jats:sec> <jats:sec> <jats:title>Growing points</jats:title> <jats:p>AI has potential to be used in virtual review clinics, telemedicine and as a training tool for junior doctors. Unsupervised AI techniques offer the potential of uncovering currently unrecognized patterns of disease. If this promise is fulfilled, AI may then be of use in challenging cases or where a second opinion is desirable.</jats:p> </jats:sec> <jats:sec> <jats:title>Areas timely for developing research</jats:title> <jats:p>There is a need to determine the external validity of deep learning algorithms and to better understand how the ‘black box’ paradigm reaches results.</jats:p> </jats:sec>

Interpreting Deep Learning Studies in Glaucoma: Unresolved Challenges

“…Besides the predominance of ethnically Chinese subjects in this dataset, the exclusive acquisition from clinical settings means this dataset may exclude patients with undetected glaucoma, therefore underrepresenting the prevalence of glaucoma in the general population. 21 Similar limitations are present in the training datasets used by Liu et al 45 (who used the Chinese Glaucoma Study Consortium (CGSC), 65 which is 78.3% Han Chinese), Asaoka et al 47 (who used a database composed of patients from 7 hospitals or clinics across Japan), and Li et al 53 in their VF study (who used data collected from 3 ophthalmic centers in China). When Liu et al 45 compared the accuracy of their CGSC-trained deep learning algorithm on a local validation dataset to its performance on a multiethnic external validation dataset (from the Hamilton Glaucoma Center), the AUC of their algorithm dropped from 0.996 to 0.923, sensitivity dropped from 96.2% to 87.7%, and specificity dropped from 97.7% to 80.8%.…”

“…Most machine learning studies in glaucoma have so far excluded subjects with multiple ocular pathologies, so it is unclear what effect other ocular pathologies can have on the accuracy of deep learning algorithms. 7,8,21 Asaoka et al 47 (SD-OCT study) and Thompson et al 49 specifically excluded subjects with ocular or other systemic diseases that could adversely affect the optic nerve or visual field from their training datasets. Thompson et al 49 excluded eyes with refractive error = þ6.0 or -6.0 diopters, whereas Asaoka et al 47 also excluded subjects with ocular disorders that could affect VF, eyes with possible secondary ocular hypertension, and eyes with anomalous discs including tilted discs.…”

“…http://www.labelme.org), which are obtained from various clinical settings in China. Besides the predominance of ethnically Chinese subjects in this dataset, the exclusive acquisition from clinical settings means this dataset may exclude patients with undetected glaucoma, therefore underrepresenting the prevalence of glaucoma in the general population 21 . Similar limitations are present in the training datasets used by Liu et al 45 (who used the Chinese Glaucoma Study Consortium (CGSC), 65 which is 78.3% Han Chinese), Asaoka et al 47 (who used a database composed of patients from 7 hospitals or clinics across Japan), and Li et al 53 in their VF study (who used data collected from 3 ophthalmic centers in China).…”

“…Several reviews have summarized the results and clinical applicability of existing deep learning studies relating to glaucoma. 7,8,[18][19][20][21] The purpose of this review is to expand on their discussion by examining the largest deep learning studies for diagnosing glaucoma (eTable 1; http://links.lww.com/APJO/ A83) and identifying recurring challenges and reasons limiting autonomous AI for glaucoma. We will cover the 3 most common deep learning inputs for glaucoma, challenges in defining primary open-angle glaucoma and in establishing ground truth to adhere to the definition, and dataset characteristics limiting algorithm generalizability.…”

Human, All Too Human? An All-Around Appraisal of the “Artificial Intelligence Revolution” in Medical Imaging

“…Likewise, also clinicians need not fear AI as a potential enemy who could harm their professional reputation in the patients’ eyes or their jobs in the future, but they should leverage its power to tackle computationally and labor-intensive tasks better than humans and to concentrate on those tasks which require human action ( Ahuja, 2019 ). Therefore, an enhanced professional role could be envisaged for both radiologists and clinicians, requiring more advanced and specific skills ( Recht and Bryan, 2017 ; Krittanawong, 2018 ; Ahuja, 2019 ; Waymel et al, 2019 ), despite fears that AI taking over professional tasks once performed by humans could, in the long run, lead to deskilling of human physicians ( Bisschops et al, 2019 ; Campbell et al, 2020 ; Panesar et al, 2020 ). AI could actually help radiologists and clinicians make the most of their own specialty knowledge and competence in a medical science of rapidly increasing complexity (where “diseases do not respect boundaries” between medical specialties and require the cooperation of multiple specialists; Deo, 2021 ), avoiding misunderstandings and “turf wars” due to poor communication and confusion regarding their specialty-specific roles in patient management, and possibly fostering the adoption of AI-augmented multidisciplinary teams (including software engineers and data scientists among participants) for clinical decision making ( Di Ieva, 2019 ; Lee and Lee, 2020 ; Martín-Noguerol et al, 2021 ).…”