Report by HPC Park at the IX Industry Forum on Telemedicine

01 / 12 / 2023

Report by HPC Park at the IX Industry Forum on Telemedicine

On December 1, 2023, employees of HPC Park visited the IX Industry Forum on Telemedicine, Digitalization of Healthcare and Investment in Medicine «Telemedforum 2023/Digital Medicine», which was held in Moscow, at the technopark «Skolkovo». The purpose of the forum was to consider topical issues of digitalization of health care and application of artificial intelligence technologies.

The founder of the company Maxim Dudkin made a presentation on improving the efficiency of diagnostics by talking about the product of our partners Care Mentor AI – a comprehensive platform for analyzing medical images using artificial intelligence. The participants of the event learned how the model was trained on our HPC Park Cloud Service platform and what results our partners achieved.

The Care Mentor AI company was given access to HPC Park’s grant resources for one month for testing machine learning task with server GPU. The partners were provided with 2 Nvidia Tesla A100 accelerators and 2 TB of storage. As a result of training X-ray (thorax) recognition models achieved quality improvement (ROC AUC - percentage of model accuracy probability) and added five new pathologies to the existing eight ones.

Maxim noted that the success story costed 0 rubles, as the client has already achieved results at the test-drive stage - they improved their product and have already received a registration certificate, since any project in the medical sector is subject to mandatory certification.

Artificial intelligence technologies in medicine.

Then, aiming to put the public in context, Maxim spoke about the stages of AI development in radiation diagnostics.

So, a multi-stage process is used to effectively identify and diagnose diseases in the creation of artificial intelligence for medical pathology analysis. Starting from the phase of medical analysis, where more than 40 pathologies on computer tomography (CT) are identified, the strategic choice is made for the most common and socially important, such as tuberculosis, pneumonia, fractures and others.

This is followed by a research collection phase, where physicians carefully supervise the classification of the studies, taking into account the previous analysis. These studies are combined and stored in appropriate folders, creating a database of tens of thousands of images for each type of pathology.

In the next phase Data Science specialists develop the AI core, which will be trained on the collected data. The method of research marking is used for training, where AI learns to determine the type of pathology.

The next steps include preparing a validation dataset, training AI on a part of the total data, and running through a validation dataset to collect metrics such as accuracy and specificity, and the area below the ROC curve (ROC AUC). It is important to admit that the process is iterative and corrections in the AI core are made at each stage until the required quality metrics are achieved.

Finally, after successful completing the learning and correction phases, the AI core is turned into a finished product. However, before starting commercialization, it is necessary to go through the procedure of obtaining a certificate for medical products. This important step ensures compliance with legislation and evidence that technical and clinical trials have been completed.

After obtaining the registration certificate the moment of transition to commercialization of the product comes. The integration of AI in X-ray diagnostics is the integration into the PACS server, which is the catalog and route of research.

New horizons in the area of medical diagnosis.

The speaker summarized the development stages of AI with a brief description of two unique series of CareMentor AI’s products, which are the results of the AI analysis.

The first series represents the visual marking of the study. In the frontal X-ray image AI isolates the zone contours for further work, determines the type of pathology and creates a new series of studies, leaving the main (basic) series intact.

The second series is an AI report where a text form indicates the pathology AI detected and its probabilities in this study.

The doctor eventually compares the results of the visual analysis with the text report and after confirming and agreeing the results, inserts the generated report into his own document. This makes it possible to quickly and accurately transmit information on pathologies detected and their degree of probability into medical records.

As a result, from the point of view of the work of medical professionals, these solutions help to optimize time costs. Contrast and brightness games are now performed automatically as AI pre-identifies areas with potentially targeted pathologies. While the time spent on imaging diagnostics increases, it is compensated by the efficiency of the system’s further application. If a doctor had previously needed to spend three minutes to analyze a healthy patient and four minutes on a patient with pathology, then after the innovations of analyzing pictures without pathologies it takes about a minute, while the patient with the abnormalities detected - the rest of the time. It is important to note that artificial intelligence acts in symbiosis with the doctor, not replacing it, but only as a tool and helping to optimize processes.

All articles