The purpose of this study was to evaluate a semi-supervised Mean Teacher (MT) framework for semantic segmentation of cataract surgical images, addressing the challenge of limited labeled data in real-world clinical applications. We adapted the MT framework for four-class segmentation of the iris, pupil, intraocular lens, and surgical instruments using a small labeled set and 40,000 unlabeled Cataract-1K frames. Performance was assessed using Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff Distance (HD95). MT was compared with a fully supervised UNet under varying labeled and unlabeled conditions, and additional baselines provided broader context. Ablation studies evaluated noise types and key hyperparameters, including consistency weight (λ) and exponential moving average (EMA) decay (α). Internal validation used Cataract-1K, and external testing was done on CaDIS and CatInstSeg. Our MT model consistently outperformed the supervised UNet baseline. With 100 labeled images and 40,000 unlabeled frames, MT achieved a DSC of 0.76 ± 0.12 compared to UNet with 0.59 ± 0.17 (t = 21.23, P < 0.05). On CaDIS and CatInstSeg, MT reached DSCs of 0.69 ± 0.17 and 0.71 ± 0.20, outperforming UNet at 0.62 ± 0.08 and 0.64 ± 0.23. Optimal performance was observed with λ = 0.1, α = 0.995, and Gaussian noise σ = 15. The MT framework provides an effective semi-supervised solution for surgical image segmentation with limited annotations. Full source code and utility scripts will be released upon acceptance at: https://github.com/mahtabfaraji1/Semi-supervised-segmentation-of-cataract-surgical-images. Accurate segmentation of ocular anatomy and instruments supports surgical guidance, intraoperative decision making, and training tools in data-limited clinical environments.

To investigate disparities in lesion areas measured using external eye photography with corneal fluorescein staining (CFS) and anterior segment optical coherence tomography (AS-OCT)-based corneal epithelial thickness (CET) mapping in recurrent corneal erosion (RCE), aiming to provide a basis for laser and surgical planning. Fifteen RCE patients underwent same-day external eye photography and AS-OCT. CFS lesions were manually delineated, and CET lesions were identified by marking regions with a thickness ≥ 60µm. The two methods were compared retrospectively. Patients with extensive CET mapping-identified lesions underwent phototherapeutic keratectomy (PTK) after manual epithelial debridement, with clinical outcomes and CET map findings evaluated. The lesion area defined by CET mapping was significantly larger than that measured by CFS, with a significant median difference of 9.26mm2 (95% confidence interval [CI], 5.87-14.42; P < 0.0001). Although the two methods were strongly correlated (r = 0.904; P < 0.0001), Bland-Altman analysis revealed a consistent mean bias. In 10 cases where CFS revealed only inferior involvement, CET mapping detected superior extension in five eyes, enhancing treatment planning accuracy. PTK was performed on nine patients, with no recurrences occurred (1 loss to follow-up). Smooth corneal epithelium with normalized thickness were observed in CET maps at a median of 7 months (range, 3-10; n = 5). CET mapping is a sensitive and reliable tool for detecting larger lesion areas in RCE compared to CFS, enabling more accurate treatment planning and improved management outcomes. AS-OCT-based corneal epithelial thickness mapping can refine clinical decision-making by revealing subclinical epithelial changes, supporting personalized treatment strategies for RCE.

PurposeTo analyze the genetic basis and phenotypic features of occult maculopathy in the Chinese population.MethodsClinical data and venous blood from probands with occult maculopathy and available family members were collected. Variants were detected using next-generation sequencing (NGS), analyzed bioinformatically, and confirmed through Sanger sequencing.ResultsHeterozygous pathogenic variants in RP1L1 were detected in 17 of 24 families, with c.133C>T and c.3599G>T identified in 15 and 2 families, respectively. In patients with the RP1L1 variants, best-corrected visual acuity (BCVA) showed an overall declining trend despite sporadic fluctuations, stabilizing between 0.70 and 1.0 logarithm of the minimum angle of resolution in most patients. No potential pathogenic variants were detected in the remaining seven families (7/24, 29.17%) after systematic analysis of NGS data. Comparative assessment revealed that patients lacking RP1L1 variants had significantly earlier onset and better BCVA at their first visit. A distinctive finding was universally normal optical coherence tomography (OCT) profiles in this non-RP1L1 group. Other atypical changes were detected in subsets of patients on other examinations, including OCT, fundus autofluorescence, microperimetry, and adaptive optics retinal imaging.ConclusionsHeterozygous pathogenic variants in RP1L1 contributed to 70.83% (17/24) of Chinese families with occult maculopathy in our cohort. Multifocal electroretinogram (mfERG) remains the most sensitive detection modality for occult maculopathy, while bilateral OCT abnormalities might indicate a genetic etiology related to heterozygous RP1L1 variants. Other modalities, including fundus autofluorescence, visual field, and adaptive optics imaging, have been explored for diagnosis and differentiation.Translational RelevanceThis study emphasizes the importance of genetic diagnosis and mfERG for the accurate diagnosis and management of occult maculopathy.

Human retinas have unique anatomical structures called the macula and an associated characteristic vascular pattern. Despite its clinical importance, the mechanism underlying the human-specific vascular pattern remains unknown because of limitations in experimental approaches using primate samples. This study aimed to elucidate such vascular formation process. We first examined the effects of four hypothetical factors contributing to foveal avascular zone (FAZ) formation: inhibitory molecule secretion, chemoattractant depletion, tissue deformation (towing), and tip cell migration restriction. None reproduced the features of the human retinal vascular structure. Next, we developed a mathematical model of human retinal vascular development by considering endothelial cells and astrocytes. We assumed that retinal vessels form via angiogenesis according to the gradient of vascular endothelial growth factor and that astrocytes dynamically expand while avoiding the fovea, providing scaffolds for angiogenesis. Our astrocyte-coupling model recapitulated various features of the human retinal vascular pattern, including a radially outward vascular pattern from the optic disc, inferior and superior temporal arcades, FAZ formation, a radially inward vascular pattern around FAZ, and a vertically facing pattern toward the horizontal vascular borderline in the temporal region of FAZ. Other models did not support the other four hypotheses. Our model explained human-specific retinal vascular formation as the combination of angiogenesis and vascular growth restriction by retinal astrocytes. These results also suggest the importance of astrocyte dynamics, particularly their timing of spreading. Our modeling framework can be extended to abnormal vascular patterns observed in diseases, including retinopathy of prematurity.

Differentiating pseudopapilledema from papilledema is challenging, but critical for prompt diagnosis and to avoid unnecessary invasive procedures. This study evaluates automated machine learning (AutoML) model performance for distinguishing the presence and severity of papilledema using near infrared reflectance images obtained from standard optical coherence tomography, comparing the performance of different AutoML platforms. . A retrospective cohort study was conducted using University Hospitals of Leicester, NHS Trust data. Optic nerve head-centered OCT imaging was obtained for 289 patients (813 images) from 2021 to 2024, with normal optic discs (69 patients, 185 images), papilledema (135 patients, 372 images), and optic disc drusen (ODD) (85 patients, 256 images). AutoML platforms-Amazon Rekognition, Medic Mind, and Google Vertex-were evaluated for (1) distinguishing papilledema from normal discs and ODD and (2) grading papilledema severity (mild/moderate vs. severe). Model performance was evaluated using area under the curve (AUC), precision, recall, F1 score, and confusion matrices for all six models. Amazon Rekognition showed the best performance in distinguishing papilledema from normal/ODD (AUC, 0.90; F1 score, 0.81) and grading severity of papilledema (AUC, 0.90; F1 score, 0.79), outperforming Google Vertex and Medic Mind, which had slightly lower accuracy and higher misclassification rates. This evaluation demonstrates the feasibility of AutoML platforms in papilledema classification using near-infrared reflectance images obtained from standard optical coherence tomography. Further external validation is needed to confirm clinical utility. Automated machine learning can be feasibly used to provide an accessible, scalable solution for clinical teams without coding expertise to recognize and characterize papilledema.

Distinguishing uveal melanoma (UM) from uveal nevus (UN) is often challenging yet crucial for appropriate management. Machine learning (ML), particularly deep learning (DL), has emerged as a promising solution to this binary classification task. This study aimed to examine the ability of artificial intelligence (AI) models to differentiate UM from UN based on fundus photographs. We systematically searched four databases through July 6, 2025, for studies developing ML models for distinguishing UM from UN using fundus photographs as input. The risk of bias and applicability concerns were assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 tool. The results of primary studies were pooled using random-effects meta-analysis. Our review included seven articles with 6208 participants in total. Six studies used DL and one applied conventional ML. Only two studies conducted external validation. The proposed algorithms demonstrated a strong performance, achieving a pooled area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity of 0.915, 85.3%, 83.7%, and 87.7%, respectively. In subgroup analysis, externally validated models retained a promising discriminative ability with a subtotal pooled AUC of 0.873. ML algorithms showcase consistently high performance in differentiating UM from UN based on fundus photographs, supporting their potential role as adjunctive tools in clinical practice. Their objective, reproducible assessments may improve referrals, guide clinical decision-making, and boost diagnostic confidence across health care providers. However, existing evidence is highly heterogeneous and constrained by small dataset sizes and limited external validation. Addressing these gaps through multicenter collaborations and data-sharing initiatives could yield more accurate, robust, and generalizable models. This work bridges computational research and ophthalmologic care by demonstrating the potential of AI-based analysis of fundus photographs to assist in differentiating uveal melanocytic tumors.

Drusen are subretinal pigment epithelium extracellular matrix deposits. This study compared histomorphometric characteristics of age-related and atypical drusen found in an eye of a young woman with chronic retinal detachment (RD). We conducted a histomorphometric examination of drusen identified in 3 eyes: an eviscerated eye from a 35-year-old woman with chronic RD, and 2 specimens of 2 patients, aged 91 and 77years, with typical age-related drusen. Sections of formalin-fixed tissue stained with hematoxylin and eosin (H&E), periodic acid-Schiff (PAS), Masson trichrome, luxol fast blue (LFB), and Von Kossa were evaluated using light microscopy. Drusen dimensions were quantified using Philips Digital Pathology Software. Atypical drusen identified in the RD eye (n=25) displayed spheroidal shape with peripherally radiating fibrils. There were foci of dystrophic calcification and localized chronic choroidal inflammation and within some drusen. Age-related drusen (n=29) had a typical dome-shaped appearance. Atypical drusen had significantly larger (P<0.001) average circumference (278 vs. 52µm), diameter (96 vs. 20µm), and area (0.006 vs. 0.00022mm2) than age-related drusen, respectively. Both groups stained positively with PAS, negatively with LFB and Von Kossa, and blue on Masson Trichrome. Atypical drusen found in this RD case display significant differences in morphometric parameters compared to age-related drusen but similar histochemical staining patterns, suggesting a similar macromolecular composition. Further investigations are needed to inform drusen pathogenesis, morphology, and associations with sex, age, or long-standing RD. This case adds to the limited body of evidence on drusen in younger patients and suggests that atypical drusen may form in association with RD.

PurposeThe purpose of this study was to evaluate the accuracy and perceptual quality of red-green-blue (RGB) fundus images generated from red-green (RG) images using Pix2Pix-based generative adversarial network (GAN).MethodsRG images were extracted from original RGB images recorded by a newer model of the Optos system by turning off the blue laser source. A conditional GAN based on the Pix2Pix architecture was trained to convert this extracted RG image to an RGB image. The performance was evaluated quantitatively by using a five-fold cross-validation and assessments with structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), learned perceptual image patch similarity (LPIPS), mean absolute error (MAE), and root mean square error (RMSE). To assess the perceptual accuracy, 39 ophthalmologists examined 60 randomly selected images, 30 true and 30 generated RGB images, using Google Form-based interface. They were instructed to classify each image as a true or an artificial intelligence (AI)-generated image.ResultsThe generated images had high similarity to the original RGB images (SSIM = 0.97 ± 0.00; PSNR = 35.74 ± 0.17 decibel [dB]; LPIPS = 0.09 ± 0.00; MAE = 1.76 ± 0.04; and RMSE = 4.25 ± 0.10). The overall correct classification rate was 57.1%. Certain lesion types, such as epiretinal membrane and optic disc cupping, had higher discrimination rates. The receiver operating characteristic (ROC) analysis across all evaluations yielded an area under the curve (AUC) of 0.497 (P = 0.314).ConclusionsPix2Pix-based GANs can generate perceptually and structurally consistent RGB images from RG images without lesion-specific attention mechanisms.Translational RelevanceThis method enables a reinterpretation of legacy RG fundus images and may support future AI diagnostic applications.

PurposeTo identify factors influencing best corrected visual acuity (BCVA) prognosis in patients with AQP4-IgG+ neuromyelitis optica spectrum disorder–related optic neuritis (NMOSD-ON) after treating with high-dose intravenous methylprednisolone.MethodsA total of 161 intravenous methylprednisolone (IVMP)–treated patients were randomized into training (n = 113) and internal validation (n = 48) sets at a ratio of 7:3. Using Least Absolute Shrinkage and Selection Operator regression, 16 candidate predictors were screened to construct a logistic model to predict follow-up BCVA after six months post-ON attack. Discrimination (area under curve [AUC]), calibration, and clinical net benefit (decision curve analysis [DCA]) were assessed.ResultsThe final model included seven predictors: age, AQP4-IgG titer, non-ON attack history, baseline BCVA, mean deviation, macular ganglion cell inner plexiform layer thickness, and treatment window. AUCs were 0.904 (95% confidence interval [CI], 0.850–0.959) and 0.864 (95% CI, 0.761–0.967) in training and validation sets, respectively. Calibration curves indicated high consistency, and DCA demonstrated significant net benefit at 10%–90% risk thresholds.ConclusionsThe validated prediction model effectively stratifies IVMP-treated patients at risk of visual disability, enabling precision management.Translational RelevanceBy converting heterogeneous prognostic factors into a practical prediction tool, this work empowers clinicians to deliver precision care for a blinding complication of NMOSD, directly reducing preventable visual disability.

Our previous studies in Royal College of Surgeons (RCS) rats demonstrated that co-administration of MER proto-oncogene tyrosine kinase (MERTK) and its ligand Gas6 effectively protected photoreceptors and improved visual function. To facilitate clinical translation, we developed Gas6-loaded sustained-release liposomes (Gas6 Lips) using liposomes with superior biocompatibility and enhanced release stability as the delivery vehicle, and subsequently evaluated their long-term efficacy and safety in RCS rats and nonhuman primates. Gas6 Lips were synthesized using a thin-film hydration method and characterized for morphology, encapsulation efficiency, and release profile. A GMP-grade adeno-associated virus serotype 2 (AAV2)-BEST1-hMERTK vector was constructed. Experimental groups included the AAV-hMERTK group, the combination group with AAV-hMERTK and Gas6 Lips, and the control group. Therapeutic efficacy was assessed in RCS rats using electroretinography (ERG) and histological analysis. Long-term safety was evaluated in cynomolgus monkeys over 12 months. Gas6 Lips had an average size of 144.1 nm and sustained Gas6 release for over 5 weeks. In RCS rats, the combination group exhibited significantly higher ERG b-wave amplitudes (52.116 ± 9.747 µV vs. 36.186 ± 7.156 µV, P < 0.01) and greater outer nuclear layer preservation (14.47 ± 1.43 µm vs. 10.41 ± 1.19 µm, P < 0.01) compared with the AAV-hMERTK group. These therapeutic effects persisted for over 7 weeks. In primates, no significant changes in retinal structure, ERG, or systemic immune responses were observed, with only transient and reversible mild ocular inflammation in the early post-treatment phase. Gas6 Lips combined with AAV2-BEST1-hMERTK effectively preserved retinal structure and function in RCS rats and demonstrated long-term safety in primates, supporting its potential for clinical translation. This can provide preclinical efficacy and safety data for the clinical translation of Gas6 Lips for the treatment of retinal pigmentosa (RP).