Grzybowski Andrzej, Peeters Freya, Barão Rafael Correia, Brona Piotr, Rommes Stef, Krzywicki Tomasz, Stalmans Ingeborg, Jacob Julie
Institute for Research in Ophthalmology, Foundation for Ophthalmology Development, Poznan, Poland.
Department of Ophthalmology, University Hospitals UZ Leuven, Leuven, Belgium.
Acta Ophthalmol. 2025 Jun;103(4):388-395. doi: 10.1111/aos.17428. Epub 2024 Dec 10.
To compare two artificial intelligence (AI)-based Automated Diabetic Retinopathy Image Assessment (ARIA) softwares in terms of concordance with specialist human graders and referable diabetic retinopathy (DR) diagnostic capacity.
Retrospective comparative study including 750 consecutive diabetes mellitus patients imaged for non-mydriatic fundus photographs. For each patient four images (45 degrees field of view) were captured, centered on the optic disc and macula. Images were manually graded for severity of DR as no DR, any DR (mild non-proliferative diabetic retinopathy [NPDR] or more), referable DR (RDR (more than mild DR)), or sight-threatening DR (severe NPDR or more severe disease and/or clinically significant diabetic macular edema [CSDME]). IDx-DR and MONA DR output was compared with manual grading and with each other.
Total sample size was 750 patients, of which 55 were excluded due to ungradable images. Out of the remaining 695 patients 522 (75%) were considered as having no DR by manual consensus grading, and 106 (15%) as having RDR. Agreement between raters varied between moderate to substantial. IDx-DR showed moderate agreement with human grading (k = 0.4285) while MONA DR had substantial agreement (k = 0.6797). Out of 106 patients with a ground truth of RDR, IDx-DR identified 105 and MONA DR identified 99. The sensitivity and specificity rates for RDR detection of IDx-DR were 99.1 and 71.5% compared with MONA DR of 93.4 and 89.3% respectively. Of note, both ARIAs had 100% sensitivity for the detection of STDR.
Both ARIAs performed well in this study population, both with sensitivity for RDR screening over 90%, with IDx-DR showing higher sensitivity and MONA DR higher specificity. MONA DR showed superior agreement with human certified graders.
比较两种基于人工智能（AI）的自动糖尿病视网膜病变图像评估（ARIA）软件与专业人工分级者的一致性以及可转诊糖尿病视网膜病变（DR）的诊断能力。
回顾性比较研究，纳入750例连续的糖尿病患者，这些患者均接受了非散瞳眼底照相。为每位患者拍摄四张图像（45度视野），以视盘和黄斑为中心。图像由人工根据DR严重程度分级为无DR、任何DR（轻度非增殖性糖尿病视网膜病变[NPDR]或更严重）、可转诊DR（RDR（超过轻度DR））或威胁视力的DR（严重NPDR或更严重疾病和/或临床显著性糖尿病黄斑水肿[CSDME]）。将IDx-DR和MONA DR的输出结果与人工分级结果以及彼此之间进行比较。
总样本量为750例患者，其中55例因图像无法分级而被排除。在其余695例患者中，经人工共识分级，522例（75%）被认为无DR，106例（15%）有RDR。评分者之间的一致性从中度到高度不等。IDx-DR与人工分级显示中度一致性（k = 0.4285），而MONA DR有高度一致性（k = 0.6797）。在106例确诊为RDR的患者中，IDx-DR识别出10例，MONA DR识别出99例。IDx-DR检测RDR的敏感性和特异性分别为99.1%和71.5%，而MONA DR分别为93.4%和89.3%。值得注意的是，两种ARIA软件检测严重威胁视力的DR（STDR）的敏感性均为100%。
两种ARIA软件在本研究人群中表现良好，筛查RDR的敏感性均超过90%，IDx-DR显示出更高的敏感性，MONA DR显示出更高的特异性。MONA DR与经人工认证的分级者显示出更好的一致性。
