Purpose: to evaluate of the potential for diagnosing ophthalmological and systemic diseases from fundus images using a multimodal transformer model trained on an open dataset.

Material and methods. An open RFMiD dataset containing 3200 fundus images annotated across 29 disease classes was used for training and validation. A pre-trained multimodal transformer architecture was used and fine-tuned on this dataset.

Results. The model demonstrated stable convergence and high accuracy in identifying 29 disease classes from fundus images, achieving a test AUC of 0.9155 without signs of overfitting.

Conclusion. The obtained results show high performance of the multimodal transformer-based model for the task of multiclass disease classification from fundus images.

Fundus imaging plays a crucial role in the diagnostics and monitoring of the diseases of the retina, optic nerve, and choroid. Over the past decades, fundus imaging technologies have evolved from simple ophthalmoscopes to high-precision tomographs and multispectral systems. This review discusses the main technologies of fundus imaging, their principles, advantages, and limitations. Modern developments are focused on increasing portability, automation, and accessibility for widespread use.

Purpose of the study — to develop a model for preoperative calculation of the optical power of intraocular lenses (IOLs) based on artificial neural networks (ANN model) with open architecture, its machine learning on local empirical data and comparison of the model error with modern fourth-generation formulas.

Materials and methods. The initial dataset included anonymized data of patients of S.N. Fedorov Tambov brahch of National medical research center “MNTK Eye Microsurgery”, and contained 890 records, including refraction of the strong and weak cornea meridians before surgery, axial length, anterior chamber depth, lens thickness, and A-constant of the IOL model used. The required optical power of the IOL was selected as the output value. To develop ANN models, standard machine learning tools of the Python language were used, as well as gradient and gradient-free methods of the author’s development, which were used in interactive mode. All technological processes were carried out in Google Colaboratory. To assess the quality of ANN models, we used the average relative error and the percentage of calculated values falling within the target range of ±0.5 D.

Results. The accuracy of fourth-generation formulas used for preoperative calculation of the optical power of IOLs — Barrett Universal II, Hill-RBF, Kane and Pearl DGS was assessed using a significant amount of local data. The average relative error is 2.67–3.21 %, the percentage of calculated values falling within the range of ±0.5 D is from 55 to 68 %. An ANN model based on machine learning has been developed, which allows calculating this indicator with an error of 2.33 %, with the percentage of calculated values falling within the target range of about 74 %.

Conclusion. The developed ANN model can be used in decision support systems for ophthalmologists in the form of a specialized calculator.

Laser vision correction (LVC) is a high-technology method of correcting refractive errors, the effectiveness of which largely depends on the precision of preoperative diagnostics, the individualization of surgical parameters, and the accurate prediction of postoperative outcomes. In recent years, machine learning (ML) algorithms have been increasingly integrated at various stages of the LVC process, contributing to the transformation of ophthalmic surgical practice toward a more personalized approach. This literature review aims to systematize current advancements in the application of ML in laser refractive surgery. In the preoperative phase, ML enables the accurate selection of suitable candidates, the identification of contraindications (including keratoconus), the choice of optimal surgical technique, and the calculation of ablation parameters. Employed algorithms, such as random forest, XGBoost, and deep neural networks, exhibit excellent sensitivity and specificity, frequently outperforming conventional clinical methods. In the postoperative period, ML is utilized to predict visual outcomes, assess the risk of refractive regression, and determine the likelihood of enhancement procedures. Despite its high predictive accuracy, the current use of artificial intelligence (AI) remains constrained by the retrospective nature of available data, the lack of large-scale clinical validation, and the ongoing need for integration into clinical workflows. This review underscores the importance of further research to establish standardized protocols, ensure interoperability, and achieve regulatory compliance for AI-based solutions in ophthalmology.

The purpose of the work is to summarize modern approaches to predicting systemic perioperative complications in ophthalmic surgery, to assess the possibilities of using preoperative risk calculators and to determine the role of clinical registries in ensuring the safety of surgical treatment.

Material and methods. The review includes domestic and foreign publications from 2020–2025, selected from the PubMed, Scopus and Web of Science databases, as well as materials from existing national and international registries (EUREQUO, IRIS, etc.). Particular attention is paid to assessing the limitations of traditional risk scales (ASA, RCRI), the role of markers of the activity of the neurovegetative system (heart rate variability — HRV and baroreflex sensitivity — BRS), as well as the potential of artificial intelligence (AI) in the development of personalized prognostic models.

Results. It was found that ophthalmic surgery, despite the low-trauma nature of the interventions, is associated with the risk of critical incidents, especially in elderly patients with a comorbid background. Known risk stratification scales do not take into account physiological predictors and are of little use when used in ophthalmic surgery. HRV and BRS have high prognostic value, but are not integrated into the models used. AI algorithms, including machine learning systems and the concept of digital twins, allow combining clinical and physiological parameters and forming personalized risk profiles.

Conclusion. The presented data confirm the need to develop specialized ophthalmic risk calculators and clinical registries that include phys iological parameters. Integration of AI into the processes of risk stratification of systemic perioperative complications and critical incidents helps to improve the safety of ophthalmic surgery in high-risk patients.

Purpose: to develop and evaluate the effectiveness of a neural network model for automatic segmentation of epiretinal membrane (ERM) in optical coherence tomography (OCT) images.

Materials and methods. The study includes 322 labeled macular OCT scans with signs of ERM: 167 from the private dataset of the “Professorskaya Plus” clinic and 155 from the public OCTDL dataset. Five architectures were selected for comparison: U-Net, Attention U-Net, TransUNet, LOCTSeg, and Tiny-UNet. Initial annotations were generated using a baseline U-Net model and underwent expert clinical validation. Segmentation performance was assessed using Dice coefficient and Intersection over Union (IoU). Annotation quality was ensured by three experienced ophthalmologists with over 10 years of clinical practice.

Results. All models demonstrated comparable Dice and IoU scores, with no statistically significant differences. Tiny-UNet showed the best balance of accuracy and computational efficiency: 570K parameters, 5× faster training per epoch than U-Net, and total training time of only 20 min. Segmentation accuracy reached Dice = 86.1 %, IoU = 78.6 %.

Conclusion. Tiny-UNet appears to be the optimal architecture for ERM segmentation tasks, offering high accuracy with minimal computational requirements. Its efficiency makes it suitable for clinical deployment, including in mobile and cloud-based telemedicine platforms.

Postkeratotomy corneal deformation (PKCD) has significant complexity in the calculation of intraocular lenses (IOL). Traditional methods of corneal topography assessment have errors, which dictates the need to develop automated solutions based on artificial intelligence (AI).

Purpose: to develop and validate a neural network model for automated analysis of corneal topographic data in order to improve the accuracy of IOL calculation in patients with PKRD.

Materials and methods. Anonymized results of medical records of 450  patients (aged 45 to 78 years) in the late period after radial keratotomy (RK) were used (95 patients underwent cataract surgery). In addition to the standard ophthalmological examination, all patients underwent Scheimpflug-imaging (Pentacam HR, Oculus, Germany). Multivariate analysis methods were carried out, a mathematical classification algorithm was developed.

Results. The developed prototype of the neural network model is able to automatically classify corneal topographic data into six types. Based on the postoperative refractive data, the predicted refractive result and correction factors for calculating the IOL using various formulas were calculated.

Conclusion. AI technologies and the correction factor database can become the basis for optimized calculation of the IOL optical power in patients with PCDR.

Participants of the III All-Russian Summit with International Participation “AIO 2025: Artificial Intelligence in Ophthalmology”, held on May 23–24, 2025 in Abrau-Dyurso (Novorossiysk) adopted a resolution summarizing its results. It was emphasized that the introduction of digital and intelligent technologies in healthcare requires systematic training of specialists. The experience of using artificial intelligence (AI) and neural network models in ophthalmology has shown that they can significantly improve the accuracy of diagnosis and prognosis of eye diseases. It is advisable to support interdisciplinary research projects and pilot software platforms for analyzing large medical data sets (Big Data), planning surgeries and modeling clinical outcomes, creating multimodal AI systems capable of complex processing of text and visual data (fundus photography, optical coherence tomography, perimetry, etc.) for the purpose of early diagnosis of eye diseases. Particular attention should be paid to the creation of AI systems for automating calculations of the optical power of lenses and assessing corneal topography in complex categories of patients. It is proposed to amend the current legislation and regulatory legal acts in terms of providing high-tech medical care and the use of intelligent technologies, including for eye diseases.