Background
Digital Vision Solutions, the developer of Retina.AI, a cloud-based ophthalmology platform, migrated its solution to Yandex Cloud. The project was deployed using the Cloud Functions serverless computing technology, Managed Service for YDB database, and the Yandex Cloud Object Storage data platform.
Using the Russian cloud allowed the company to localize the project, comply with legislative requirements for the storage of personal data, and ensure high performance of ML model scripts and uninterrupted, round-the-clock operation of the platform.
The service is currently being tested, and today it is successfully in use by more than 800 of the company’s customers, creating a load of up to 200 daily requests for image processing.
Localize the service and resolve performance and scaling issues
Digital Vision Solutions is a developer of the Retina.AI cloud-based ophthalmology platform. The company’s solution is a cloud-based service for searching for human retinal pathologies in images taken using specialized cameras that take a picture of the fundus, or optical coherence tomography devices. Retina.AI is included in the Russian Software Registry, patented, and has successfully passed clinical trials of the Federal Scientific and Clinical Center for Specialized Types of Medical Care and Medical Technologies, within the Russian Federal Medical-Biological Agency.
The service is currently being tested, and is currently in use by more than 800 users in Russia and other countries around the world.
Artificial intelligence algorithms diagnose the most common ophthalmological diseases of the human retina, including diabetic retinopathy and a number of others. The data model is trained on the results of analyzing other images and the specialized doctors’ logic of decision-making. A flexible and productive computing platform is required for the fast and error-free operation of the system.
The physical infrastructure limited the possibilities of scaling and performance, so the company decided to switch to a cloud platform. Digital Vision Solutions had experience interacting with Microsoft Azure, but deploying there would neither have allowed the solution to receive the necessary registration of a medical device, nor fulfill the requirement of Russian Federal Law 152-FZ for the storage of personal data within the territory of the Russian Federation.
In order to solve localization issues, comply with the requirements of Russian legislation, and ensure high performance of the AI system, the company decided to migrate the project to the Yandex Cloud platform.
The company chose a cloud platform based on the experience of working with similar solutions from other vendors, as well as taking into account the availability of the cloud services needed.
They made the choice in favor of Yandex Cloud after analyzing similar platforms. The decisive factors in making the decision were:
- the availability of serverless computing
- convenient managed services
- localization on Russian territory
- compliance with Russian Federal law 152-FZ
- a responsive support service
- the ability to scale resources quickly and easily
- a transparent financial model and payment only for resources consumed.
Yandex Cloud offers a high level of fault tolerance and reliability through the use of modern, geo-distributed infrastructure, and meets the strict requirements of information security and Russian legislation. Another important criterion was the high economic efficiency of using Yandex Cloud resources.
As Retina.AI is a highly specialized solution, and its developer is essentially a startup, the ability to pay only for consumed resources has allowed us to reduce computing costs and use more funds for product and business development.
Serverless computing and model optimization
The Retina.AI platform has become the company’s third project using ML technology to analyze graphical data. The first version of the models and handler scripts was hosted on a local server. This option was suitable for prototyping, but not for commercial operation, as it did not meet requirements for fault tolerance and scaling.
The company decided to migrate the platform to the cloud and use serverless computing to reduce costs, both in terms of finances and in time. In order for the platform to be designated in the government registration as a medical device, and to comply with the requirements of Russian Federal Law 152-FZ and 353-FZ regarding the storage of personal data and medical secrecy, Retina.AI migrated to Yandex Cloud.
The main elements of the platform and the services used:
- Yandex Cloud Functions — storage of the model and handler scripts
- Yandex Managed Service for YDB — a database for storing service data (snapshots, user data, and other information)
- Yandex Object Storage — storage of snapshots donated by users or with reported processing errors
- The client’s website and personal account are stored on a third-party hosting service, take snapshots from the user, pre‑process images from fundus cameras for more accurate processing, upload the images to the cloud, and return the final result to the user.
The migration of the project to Yandex Cloud took 2.5 months and was fully implemented by the company’s specialists. Technical difficulties in adapting scripts arose during the migration process, and we were able to successfully solve them together with the Yandex Cloud technical support service.
Now a snapshot takes 2-3 seconds to process, and the finished result is delivered to the client within 10 seconds. It was possible to achieve such high performance by optimizing the model (its size does not exceed several dozen megabytes) and using serverless computing. Functions used are divided into categories: snapshot processing, error analysis, admin interface, user accounting, and others. Image segmentation is performed by an ultra‑precise neural network with U-Net architecture. The images that platform users have allowed to use (donated) to train the model are stored in a reliable, secure repository.
The system currently processes up to 150-200 customer requests per day. The system is stable, and the technological stack used has great potential to scale the project.
Results
Digital Vision Solutions was able to solve the main task: hosting Retina.AI in a secure and reliable Russian cloud that ensures round-the-clock uninterrupted operation of the platform.
Using Yandex Cloud made it possible to comply with Russian legislative requirements regarding the storage of personal data and medical secrets. Using Cloud Functions has allowed us to reduce the labor costs of infrastructure maintenance and increased financial efficiency. Payment for capacities is made according to their consumption, which makes the project economically transparent.
Migrating Retina.AI to the cloud makes it possible to process 150-200 customer requests per day without any loss in performance. Cloud services are ready-made products that are easy to implement even for a small company with a few of its own technical specialists.
The company’s future plans include state registration of the service as a medical device, the introduction of the Yandex DataSphere full‑cycle ML development environment, and the transfer of the project’s frontend to Cloud Functions.
Opinion
Our company, Digital Vision Solutions, develops the Retina.AI cloud ophthalmology service project, which uses the most advanced AI algorithms to diagnose a wide range of ophthalmological diseases. In Yandex Cloud, we have found a reliable partner for our cloud service, which already employs almost 800 users from Russia, the CIS countries, and almost 25 other countries around the world. We have high requirements for a cloud solution — reliability, security, and the ability to quickly scale to thousands and tens of thousands of users. Among our competitors is Google Deep Mind and several American and European companies, so it’s important for us to focus on the product side. Yandex Cloud allows us to not have to think about the problems of organizing our service in the cloud and devote time to creatinga world-class product in Russia.