Artificial Intelligence-Based Integration Of Image And Clinical Metrics To Support Cataract Surgery Decision Making
Published 2026 - 44th Congress of the ESCRS
Reference: POCAT318 | Type: Free paper | DOI: 10.82333/x76n-tz08
Authors: Dr Yeo Kyoung Won* 1 , Ms Soon Jung Oh 2 , Ms Seung-hye An 3 , Prof. Dr Hyun Sun Jeon 4 , Prof. Dr Jisang Han 5 , Prof. Dr Myung Jin Chung 6 , Prof. Hakje Yoo 7 , Prof. Dr Dong Hui Lim 8
1Ophthalmology,Samsung Medical Center,Seoul,Korea, Republic Of;Ophthalmology,SNU Eye Clinic,Seoul,Korea, Republic Of, 2Medical AI Research Center,Samsung Medical Center,Seoul,Korea, Republic Of, 3School of Medicine,Sungkyunkwan University,Seoul,Korea, Republic Of, 4Ophthalmology,Seoul National University Bundang Hospital,Seongnam,Korea, Republic Of, 5Ophthalmology,Kangbuk Samsung Hospital,Seoul,Korea, Republic Of, 6Radiology,Samsung Medical Center,Seoul,Korea, Republic Of;Medical AI Research Center,Samsung Medical Center,Seoul,Korea, Republic Of, 7Medical AI Research Center,Samsung Medical Center,Seoul,Korea, Republic Of;Data Convergence and Future Medicine,Sungkyunkwan University,Suwon,Korea, Republic Of;Bio-Mechatronic Engineering,Sungkyunkwan University,Suwon,Korea, Republic Of, 8Ophthalmology,Samsung Medical Center,Seoul,Korea, Republic Of;Samsung Advanced Institute for Health Sciences & Technology,Sungkyunkwan University,Seoul,Korea, Republic Of;School of Medicine,Sungkyunkwan University,Seoul,Korea, Republic Of
Purpose
Cataract is a leading cause of visual impairment worldwide. The decision to perform cataract surgery is multifactorial, requiring consideration of lens opacity, patient-reported symptoms, and visual acuity. However, surgical indications remain largely subjective without a standardized quantitative framework. We aimed to develop an artificial intelligence-based classification model to identify clinically significant cataract (Lens Opacities Classification System III [LOCS III] ≥3) by integrating anterior segment images with visual acuity and optical quality metrics across four subtypes: nuclear opalescence (NO), nuclear color (NC), cortical opacity (CO), and posterior subcapsular cataract (PSC) to support cataract surgery decision-making.
Setting
Department of Ophthalmology, Samsung Medical Center (SMC), Sungkyunkwan University School of Medicine, Seoul, Republic of Korea
Methods
This retrospective study included 183 patients (275 eyes) from SMC between October 2014 and March 2019. The dataset consisted of anterior segment images, visual acuity, and optical quality metrics. The model was developed to classify four cataract subtypes (NO, NC, CO, and PSC) into clinically significant (LOCS III ≥3) and non-significant (<3) groups. Three input configurations —image-only, clinical metrics-only, and combined inputs—were evaluated to assess the contribution of each modality and the effect of multimodal integration. External validation was conducted using 188 patients with NO from Seoul National University Bundang Hospital between March and October 2023.
Results
Across all cataract subtypes, the image-only model achieved the highest performance among single-modality approaches, with an area under the receiver operating characteristic curve (AUROC) ranging from 0.714 to 0.884. Models integrating image and clinical metrics further enhanced classification performance. Although the optimal input combination varied by subtype, the addition of vision-based clinical metrics to anterior segment images yielded the largest performance gain, increasing AUROC by 0.07-0.15 on average. In external validation, the image-vision combined models confirmed stable performance, achieving AUROC values of 0.837–0.856.
Conclusions
Multimodal integration of image and clinical data improved the classification of clinically significant cataract. Incorporating vision-based clinical metrics may provide complementary information beyond image alone and help inform surgical decision-making.