Artificial intelligence-enabled non-invasive cataract diagnosis and grading system using anterior segment images

To the Editor: Cataracts, the leading cause of global visual impairment, affect nearly 80% of people aged 60–89 years and over 90% of those over 90 years old in China, with prevalence expected to reach 200 million by 2025.[1] Traditional diagnosis relies on slit-lamp examinations,[2] but inter-observer variability and limited access in low-resource settings highlight the need for improved diagnostic approaches. Artificial intelligence (AI) and deep learning, particularly convoluted neural network (CNN)-based models such as ResNet and VGG, have demonstrated strong potential for cataract grading using anterior segment images, which allow direct visualization of the lens compared with fundus photography.[3] Key limitations include variability in disease progression, demographic biases, and dependence on pupil dilation. We propose a novel dual-input deep learning model using slit-lamp and diffuse light images to enhance diagnostic accuracy, incorporate bilateral asymmetry, and support early detection across diverse clinical settings, improving patient outcomes and care accessibility. The study design is summarized in Figure 1A. Anterior segment images were collected from the First Affiliated Hospital of Wenzhou Medical University. Ethics approval was obtained from the Ethics Committee of The Eye Hospital of Wenzhou Medical University (approval number: 2024-048-K-43), and all analyses used de-identified data. From 19,126 patients who visited the Wenzhou Medical University Eye and Optometry Hospital between 2017–2021 and in 2024, we obtained 29,166 monocular samples (diffuse and slit-lamp images). After excluding 25,869 samples due to unsuitable grading quality, the final dataset comprised 3239 samples from 2420 patients. Examples of the images that were deemed unsuitable for grading were shown in Supplementary Figure 1, https://links.lww.com/CM9/C817. In total, 6478 images from these 2420 patients were used for model development. Participant demographics and clinical details were presented in Supplementary Table 1, https://links.lww.com/CM9/C817. Given the dataset’s limited size, we used five-fold cross-validation to enhance model robustness and generalization. In each iteration, four folds served as the training subset (80%), with the remaining fold for validation (20%).Figure 1: (A) Proposed AI framework for cataract classification. A total of 6478 anterior segment images were included, comprising 1746 slit and diffuse images from dilated pupils and 1493 slit and diffuse images from natural pupils. Image features extracted separately from slit and diffuse images were combined with pupil status (dilated or natural) for model development using multiple classic CNNs. (B) Receiver operating characteristic curve (left) and confusion matrix (right) for the external dataset. AI: Artificial intelligence; AUC: Area under the curve; CI: Confidence interval; CNN: Convoluted neural network.A multi-tiered grading approach involved two levels of graders with increasing expertise to validate image labels. Initially, radiologists with 5 to 15 years of experience excluded images with insufficient grading quality, such as those with blurring, misalignment, or corneal/conjunctival obstructions. A second tier of five senior radiologists, each with at least 25 years of clinical experience, reviewed a randomly selected 10% subset for label accuracy verification. This consensus grading provided a robust foundation for developing a reliable deep-learning algorithm for image-based classification. Traditional grading systems such as Lens opacities classification system II[4] require pupil dilation during slit-lamp examination, which poses risks including acute angle-closure glaucoma or allergic reactions, thereby limiting their suitability for large-scale screening. To address these constraints, we developed a custom grading standard inspired by LOCS II[4] and Emery[5] nuclear hardness classification. Severity was defined as mild (no cortical opacity in the pupil zone, nuclear hardness grade I–II, posterior capsule opacity <3%), moderate, or severe. A visual axis of 2.5 mm was selected based on the typical adult pupil diameter in bright light (2.54 ±0.45 mm), to ensure consistent, reproducible grading under non-dilated conditions. To improve model generalizability and robustness to unseen data, we applied multiple data augmentation techniques, including random cropping and resizing, random horizontal flipping, and color jittering. Our dual-input architecture incorporated both slit-lamp and diffuse images, using classic CNNs as backbones for classification, including ResNet50,[6] DenseNet121,[7] VGG19,[8] InceptionV3,[9] and EfficientNet-B0.[10] Two separate networks processed the slit-lamp and diffuse images, with fully connected layers replaced by identity layers to retain feature extraction only. We hypothesized that pupil status might influence the AI’s decision-making by providing distinct information. Therefore, we concatenated the pupil status (dilated or not) with image features extracted separately from slit-lamp and diffuse images. The combined feature vector passed through a fully connected layer with ReLU activation, with the model outputting the class with the highest probability for three-class decisions. The model was trained on batches of 32 images resized to 512 × 512 pixels, using cross-entropy loss with an Adam optimizer (learning rate of 1e–4) and a ReduceLROnPlateau learning rate scheduler, combined with mixed-precision training (using automatic mixed precision [AMP]) for efficiency. Training, validation, and testing were implemented in PyTorch (version 1.13.1, https://pytorch.org/), with data analysis conducted in scikit-learn (version 1.3.2, https://scikit-learn.org/). Graphs were generated using matplotlib (version 3.6.2, https://matplotlib.org/) and seaborn (version 0.13.2, https://seaborn.pydata.org/) to visualize performance metrics. Five advanced deep learning models (ResNet50, DenseNet121, VGG19, InceptionV3, and EfficientNet-B0) were trained for cataract classification. For internal validation, 648 images from 308 patients were tested, comprising 94 mild, 110 moderate, and 104 severe cataract cases. Model performance details are provided in Supplementary Table 2, https://links.lww.com/CM9/C817. DenseNet121 demonstrated the strongest performance, achieving an area under the curve (AUC) of 1.00 (95% confidence interval [CI]: 0.99–1.00) for mild cataract, 0.94 (95% CI: 0.91–0.96) for moderate cataract, and 0.96 (95% CI: 0.94–0.98) for severe cataract [Supplementary Figure 2, https://links.lww.com/CM9/C817]. To monitor training stability and convergence, we visualized the loss curves for the DenseNet121 model across epochs. Supplementary Figure 3, https://links.lww.com/CM9/C817 illustrates the training and validation loss curves, demonstrating a steady decrease in both losses without significant divergence, indicative of effective optimization and limited overfitting under the Adam optimizer and Reduce LROnPlateau scheduler. Evaluating accessibility and applicability of diagnostic tools in real-world settings is essential. Dilated pupil imaging improves visualization of ocular structures but requires pharmacological dilation, which may be uncomfortable for patients, time-consuming, and impractical in routine screening or remote settings. Therefore, we assessed model performance on non-dilated pupil images [Supplementary Figure 4, https://links.lww.com/CM9/C817]. To validate the efficacy of the dual-input architecture, we compared its performance against ablated models relying solely on slit-lamp images or diffuse light images. Supplementary Figure 5, https://links.lww.com/CM9/C817 presents the receiver operating characteristic (ROC) curves for these models. The dual-input model achieved an AUC of 0.95, outperforming the slit-lamp-only model (0.94) and the diffuse-only model (0.90), confirming the complementary value of integrating both modalities. To enhance interpretability, we applied integrated gradients to generate saliency maps that highlight regions most influential in the model’s predictions. Supplementary Figure 6, https://links.lww.com/CM9/C817 shows representative images with their saliency maps. These maps reveal that the AI model focuses on regions surrounding the pupil and other areas with visible abnormalities, aligning with ophthalmologists’ diagnostic focus and supporting the model’s clinical relevance. To further evaluate clinical applicability, we tested the model on an external cohort of 541 anterior eye segment images (187 mild, 205 moderate, and 149 severe cataracts) from the Hangzhou branch of Wenzhou Medical University Eye Hospital [Supplementary Figure 7, https://links.lww.com/CM9/C817]. The DenseNet121-based cataract grading model demonstrated robust overall diagnostic accuracy (88.7%) and maintained balanced performance across severity classes, achieving a weighted average F1-score of 0.89. As shown in Figure 1B, the model achieved an AUC of 0.98 for mild cataract, 0.94 for moderate cataract, and 0.98 for severe cataract. These results demonstrated strong generalizability to unseen clinical data and support potential real-world implementation. To assess the clinical alignment with LOCS II, we conducted a consistency analysis using Cohen’s Kappa coefficient in the external cohort. Because LOCS II provides detailed grading of cortical (C), nuclear (N), and posterior subcapsular (P) opacities, we mapped its categories to our three severity levels: mild (C0, Control, C1; N0, N1; posterior subcapsular opacity with vertical height <3% [P0, <P1]); moderate (C1 <C ≤C3; N2; posterior subcapsular opacity <30% [<P2]); and severe (C >C3, N3, or posterior subcapsular opacity ≥30% [≥P2]). This mapping ensured compatibility between LOCS II’s granular assessments and our simplified framework, which prioritizes applicability in non-dilated settings. The analysis yielded a Cohen’s Kappa coefficient of 0.85, indicating substantial agreement beyond chance. This robust concordance validates the model’s reliability as a surrogate for expert LOCS II assessments, despite minor discrepancies potentially arising from the simplified mapping or reduced visibility of subtle opacities in non-dilated images (consistency rate of 72.23% where grades matched exactly across pupil conditions). These findings demonstrate the model’s reliability as a surrogate for expert LOCS II grading and its suitability for large-scale screening. In conclusion, this study hypothesized that an AI model could identify cataract stages with accuracy comparable to experienced ophthalmologists or detect subtle, subclinical changes undetectable to the human eye. The model performed strongly on both dilated and non-dilated pupil images, supporting its applicability in real-world settings, particularly in situations where pupil dilation is not feasible or desirable. These findings suggest that this AI tool could serve as a non-invasive, high-throughput, and cost-effective screening tool for routine health assessments, particularly in general and community healthcare settings. However, further validation in larger and more diverse populations is necessary to assess broader applicability. Incorporating longitudinal data may also enable disease monitoring and predictive modeling for personalized treatment planning. Funding The study was supported by grants from the Guangzhou National Laboratory (No. YW-SLJC0201) and the Macau Science and Technology Development Fund, Macao (Nos. 0007/2020/AFJ, 0070/2020/A2, and 0003/2021/AKP). Conflicts of interest None.