Cataract is the leading cause of blindness, with ~12.6 million cases of cataract worldwide (1). The visual impairment caused by cataract can be treated with advanced cataract surgery, which can ensure a progressively better quality of vision and fewer complications than in the past (2–5). The most important indications for cataract surgery are preoperative visual acuity and cataract grading, and advances in surgical technology have expanded the scope of the surgery to even include less severe cataracts (5).

Although the cataract grading system shows a good correlation with surgical difficulty, indicating that the surgery becomes more challenging as the cataract grading increases (6, 7), it shows limitations in reflecting the patient's visual symptoms, especially in cases with nuclear cataract and cortical opacity (8). Cataract grading depends on the subjective competence of the investigator (9). However, visual acuity reflects the patient's symptoms and influences surgical difficulty; therefore, the surgery becomes more challenging also as the visual acuity decreases (10, 11). Moreover, preoperative visual acuity can serve as an important predictor of postoperative vision in various diseases (12–14).

Artificial intelligence (AI) is being increasingly used in medicine, and ophthalmology is one of the most active fields for its clinical application (15). Recent studies have attempted to use AI for cataract grading with various methods, including slit-lamp photography, fundus photography, and optical coherence tomography (OCT) (16–22), and the results suggest that AI-based cataract grading shows acceptable performance with 70–90% accuracy. However, an AI-based approach for evaluation of visual acuity in patients with cataracts is still lacking. An approach linking objective image data with the subjective symptoms represented by visual acuity is particularly relevant, since the resultant method would encompass both visual acuity and cataract grade, which are the main indications for cataract surgery. Therefore, we attempted to implement an AI model that can evaluate cataract severity based on visual acuity by using multi-source OCT data and to assess the applicability of this AI model in actual clinical practice.

The study was conducted at the Department of Ophthalmology, Severance Hospital, Yonsei University College of Medicine in accordance with the ethical standards of the Declaration of Helsinki, and institutional review board approval was obtained for the study protocol (4-2021-1697). The institutional review boards waived the need for informed consent because of the retrospective and de-identified nature of the study.

Table 1 shows the overall demographics and characteristics of the training/validation and test datasets. Although the OCT image source and underlying disease differed between the datasets, no remarkable differences were observed in age, sex, preoperative BCVA, axial length, lens thickness, and LOCS III cataract grade. The postoperative BCVA differed between the datasets (p < 0.001).

In the test dataset, the MAE, RMSE, and MPE of the model performance were 0.050 ± 0.130 logMAR, 0.140 ± 0.134 logMAR, and 1.3 ± 13.9%, respectively (Table 2). The correlation coefficient (R) between the true preoperative BCVA and estimated preoperative BCVA was 0.969 (p < 0.001) (Figure 2). The percentage of cases in the AE ≥ 0.1 group was 8.4%, and 1.9% had AE ≥ 0.2.

The preoperative BCVA values were significantly different between the AE_≥0.1 and AE _<0.1 groups (0.30 ± 0.08 vs. 0.12 ± 0.10, respectively; p < 0.001), and corneal astigmatism, lens thickness, and nucleus cataract in the LOCS III grading were significantly different (p < 0.001, < 0.001, and 0.024, respectively) (Table 3). Postoperative BCVA in the AE_≥0.1 group was worse than that in the AE _<0.1 group (p < 0.001). In the AE_≥0.1 group, the incidence of postoperative BCVA > 0.1 was 21.4%, which was higher than that in the AE _<0.1 group, and the proportion of vision-impairing disease was also higher than in the AE _<0.1 group. The percentage of cases with vision-impairing diseases was significantly higher in the AE_≥0.1 group than in the AE _<0.1 group (54.8 vs. 12.9%; p < 0.001).

The AE_≥0.1 group was further divided into underestimation and overestimation groups (Table 4). Pre- and postoperative BCVAs were significantly worse in the underestimation group (0.39 ± 0.10 vs. 0.15 ± 0.09 in overestimation, and 0.10 ± 0.15 vs. 0.01 ± 0.05; p < 0.001 and p < 0.001, respectively). The proportion of cases with postoperative BCVA > 0.1 was higher in the underestimation group (p < 0.001). The proportion of cases with vision-impairing disease was 95.7% in the underestimation group. Axial length was longer, and cortical opacity and posterior subcapsular opacity were more severe in the underestimation group (p < 0.001, 0.045, and 0.002, respectively).

Error analysis for model performance was conducted by performing comparative analyses between the AE_≥0.1 and AE _<0.1 groups and between the underestimation and overestimation subgroups of the AE_≥0.1 group (Table 5). Preoperative corneal astigmatism, lens thickness, and nuclear opacity in the LOCS III grading were higher in the AE_≥0.1 group. Increased BCVA and vision-impairing disease were more significant in the underestimation subgroup. Axial length was longer, and the cortical/posterior subcapsular opacity was more severe in the overestimation subgroup than in the underestimation subgroup.

In the AI modeling process, reducing the original image decreased the model performance, and the neural network could not estimate the BCVA when the area was reduced by ~≥90% (32% of width and height) (Figure 3A). Image augmentation increased the model performance in terms of the MAE (Figure 3B). The ensemble model had the lowest MAE, and Inception-v4 and ResNet were almost similar in terms of MAE (Figure 3C).

With the test dataset of this study, over 90% of cases could be estimated in their BCVA under 0.1 logMAR of AE. Most of the underestimation errors were caused by vision-impairing disease, and about half of the decreased model performance could be explained by a clinical approach. The technical issues, image resolution, diversity of image forms in the training dataset, and model architecture also affected the model performance.

Objective clinical assessment for estimation of the subjective visual symptoms represented by visual acuity, which is considered the ultimate goal of ophthalmologic interventions, can be utilized in various ways from clinical to experimental; however, this is very difficult for clinicians due to confounding factors (27). AI is expected to help human evaluators perform difficult tasks (28). This study suggested that cataract severity can be quantified as visual acuity via AI using OCT images.

In this study, cross-instrument validation was performed using different combinations of anterior segment OCT image sources for the training/validation dataset and the test dataset. To enhance the model performance, the OCT images used in the training/validation dataset were obtained from three devices with different detection methods, image resolutions, and image directions. The test dataset was constructed to evaluate the practical application of the AI model in real-world scenarios by including all patients within a certain period and using commonly available OCT images in the process of cataract surgery and its preparation.

Model performance and error analysis were conducted from both clinical and engineering perspectives. Although medical issues have rarely been mentioned in previous AI research, for clinical application of medical AI, error analysis from the clinical perspective is indispensable (29, 30). The results suggest that the predisposing disease and the conditions, such as high values of corneal astigmatism, axial length, lens thickness, and severity of each cataract subtype, which are well known to cause visual impairment, were the main factors underlying errors in the medical approach (8, 31–34). For engineering issues, this study suggested that low-resolution images led to the degradation of the model performance (see Figure 3A), and highlighted the importance of high-resolution images in analyzing precise medical observations (35). Multi-source images were used, and the transformation for direction and size through image augmentation in the training dataset was shown to yield better model performance (36). Although the AI model using a unified material may show a different performance from that using multi-sources materials, we focused on the versatility of the AI model by using OCT images acquired with multiple instruments.

The primary goal of cataract surgery is to restore the best possible vision as well as remove the natural crystalline lens (37). Thus, postoperative visual acuity is the target indicator of cataract surgery. Wei et al. reported the use of AI with fundus OCT for predicting postoperative visual acuity in patients with high myopia (38), and the result showed the lowest MAE of 0.16 logMAR and RMSE of 0.24 logMAR. However, this performance level was still insufficient for clinical application, and the findings implied that visual acuity was not determined by a single defined etiology. Clinicians frequently encounter patients with visual impairment without a specific pathologic lesion. This phenomenon may be caused by medical issues and unknown predisposing factors, including developmental problems, such as amblyopia, an extraocular disease, such as brain lesion, or a temporary problem, such as dry eye disease. Thus, approaches based on a single etiological factor showed limited ability to predict the postoperative BCVA. Since multiple factors, including the presence of both cataract and comorbid disease, can influence the accuracy of prediction of postoperative BCVA with AI based on the preoperative BCVA, our study could serve as the basis for future studies. Future studies should aim to predict postoperative visual acuity using a combination of preoperative visual acuity, anterior segment OCT, and other examinations.

Similar to previous medical AI research (16, 39–41), this study did not completely overcome the limitations in interpreting the neural network. In this study, the attention areas of the AI model were analyzed using heatmap analysis (Figure 4), and we tried to analyze the relationships between the attention areas and the error analysis in the medical approach. Although model training was conducted using patients with only a clinical diagnosis of cataract, the AI also recognized the cornea that was laid on the visual axis and the angle in the anterior chamber. Thus, some factors affecting the corneal shape (e.g., corneal refractive surgery), anterior chamber angle (e.g., glaucoma), and lens thickness may be related.

The strengths of this study were not limited to the excellent performance of the AI model. Our attempt to prove the importance of clinical approach in model performance was successful and our findings suggest that the addition of various clinical information in AI modeling is crucial for improving model performance.

In conclusion, the AI developed using OCT images from multiple sources showed excellent performance in estimating visual acuity in patients with senile cataracts. This quantification method encompasses both visual acuity and cataract severity of the OCT images, which are the main indications for cataract surgery, and has the potential to allow objective evaluation of cataract severity. This AI model can be used when it is difficult to express or measure the subjective visual acuity due to various causes, such as an inability to communicate. Additionally, we would like to emphasize that this was a preliminary study to expand the prediction of visual acuity after cataract surgery in patients with other diseases, possibly accompanied by visual impairment.

The datasets presented in this article are not readily available because of privacy and ethical concerns, neither the data nor the source of the data can be made available. Requests to access the datasets should be directed to HA, overhyun31@gmail.com.

The studies involving human participants were reviewed and approved by the Severance Hospital Clinical Research Ethics Committee. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

HA and T-iK: conceptualization. HA: methodology, software, formal analysis, investigation, writing—original draft preparation, and visualization. IJ, KS, EK, and T-iK: validation. IJ, KS, and T-iK: resources. T-iK: writing—review and editing and supervision. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.