Atrificial Intelligence in medicine: Expectations vs. Reality

Original article was published on Artificial Intelligence on Medium


Atrificial Intelligence in medicine: Expectations vs. Reality

All over the world developers are making creating AI services for healthcare, yet cases of real implementation of such solutions in clinical practice are in the minority. About why developers of medical AI services cannot repeat the success of colleagues in other industries and how things are going in Russia — in the material of LLC Celsus company.

Sourse: 1cdu.ru

Pretty much every leading IT corporation works on developing of smart medical products. According to the data of the research company Venture Scanner, in total, there are more than 2 thousand companies around the world working on such developments. Artificial intelligence (AI) is one of the most promising technologies in medtech. AI services can:

  • improve diagnostic accuracy;
  • automate the doctor’s work;
  • select the optimal treatment method;
  • create new pharma drugs,ect.

Probably the largest and most most talked about project on the use of AI in medicine is the American Corporation IBM and its cognitive IBM Watson system, which helps to make an accurate diagnosis and find best way to cure decease of each patient.Not so long ago, Microsoft launched the AI for Health program, under which company invests $40 million in artificial intelligence technologies for the health care.

Upcoming trend of application of AI lies in the analysis of medical images (cathodagraphy, MRI, ultrasound investigation, etc.). System is trained to identify different diseases and pathologies. In that process, technologies have achieved apparent success and therefore are already being gradually advanced to the clinic.

As for Russia, about ten companies are working on the same kind of development, including us — artificial intelligence Celsus. Currently, the system can already analyze mammograms and fluorograms, there is a development of solutions for CT and morphology.But we see the development of the product in other fields of functional diagnostics.

Next, we will analyze the main obstacles to mass introduction of AI solutions into the healthcare system and tell you about our experience in overcoming them. Some of them concern only Russia but most of the identified problems are relevant for developers from across the globe.

Neural network training: lack of datasets and universal equipment

Here one cannot do without cooperation with health care institutions and specialized professionals. In the Celsus several radiologists marking each image at once, and in case of confrontation of opinion, the images are given to roll over study.

Another technical problem is related to difference between equipment which is found in health care institutions. Images from different radiological apparatuses can visually differ greatly. If the neural network was not trained on such images, there is nothing surprising in the fact that it “fails” and does not identify anything.

At first sight, the simplest solution — is to collect all types of images from all apparatuses with all settings, mark them up and train the system. Disadvantages are that it is long and expensive. In addition, collecting all the existing types of images in the world basically seems to be an impossible task.

The best option will be usage of universal preprocessing, that is, special data handling before its delivery to the neural network. The preprocessing procedure can include automatic changes in contrast and brightness, various statistical normalization and removal of unnecessary parts of the image (artefacts).

After months of experiments, our team was able to create a universal preprocessing for radiographic images that brings practically any introductory images to a uniform appearance. This allows the neural network to process them properly.

One more feature is related to the specifics of deep learning and is relevant for many development teams in the medtech market. Medical neural networks most commonly are “heavy”, each experiment on the model takes a long time and requires huge computing powers, and therefore expensive equipment. “Celsus” has its own server on which it is possible to do four experiments parallely, as well as a cloud infrastructure which helps us to increase the number of experiments if necessary. By the way, there is a lack high-quality machine learning professionals, and they are also very expensive.

Investor and medical community confidence

There is clearly interest in medtech startups, but investors and funds are now following a very cautious strategy when it comes to medtech, realizing the difficulty of entrance to this market. Medtech has a rather long cycle of companies reaching self-sufficiency, including the need to conduct a complex and long registration procedure for a medical product.

Additionally, over the past two years, the market has been swept by a series of major bankruptcies and fraudulent activities, which could not be passed over. The most notable case is the Theranos “soap bubble”. Their blood test technology was valued at $9 billion. An equally high-profile case is related to the bankruptcy of Jawbone, a manufacturer of wearable electronics and fitness trackers. A number of respectable investment funds, such as Sequoia and Khosla Ventures, have invested more than €1 billion in the company.

A report which was published last year by London-based venture MMC company said that almost half of all European startups position themselves as related to artificial intelligence do not actually use this technology.

Overblown hype around AI justifiably causes skepticism in the medical community. Many solutions are created in isolation from the understanding of the doctor’s workflow which does not give credibility also. Developers do not have any medical education or experience working with healthcare organizations.

The decision option is to cooperate with doctors in the framework of pilot projects in health care institutions. This allows you to get across with doctors — they mark up images according to pre-approved rules, conduct tests of the system, give quick feedback, and consult developers.

A special pain in the neck for health care institutions is the comparison of different systems of competing companies, because any metrics strongly depend on the data on which the system was tested, on the percentage of patients with pathology from the total number of studies, preprocessing, and other factors. Therefore, any companies statements about the unprecedented accuracy of their models should be taken with a healthy degree of skepticism — there is no guarantee that the accuracy will not fail on new data.

To solve this problem, it is necessary to create “Golden datasets” with the help of which it will be possible to compare competing services properly and fairly. Work in this direction, including in Russia. For example, Diagnostics and Telemedicine Moscow Center actively contributes to the creation of these” Golden datasets “ in the fields of mammography, computed tomography of the lungs, and fluorography.

Legal aspects and standardization

Specificity of Russian health care, as well as a number of other countries, is that it is mostly public. Private clinics represent only a small part of the market. This fact should be taken into account by companies, because the public health service involves funding through a system of tenders and/or grants.

To use the technology in ‘real-life’ clinical practice, a certificate of registration of medical product is required. The task is complicated by the fact that the legislation of many countries (including Russia) still does not have certain standards governing the operation of medical AI services. This refers to the preparation of datasets, the performing of clinical and technical tests of services, and the standard for integrating these services into the business processes of health care institutions.

However, in many countries, the situation is changing for the better, and this is supported by the government. For example, in Russia, the government has launched the national program “Digital economy”, one of the priorities of which is the healthcare digitalization.

Vast experiment is being conducted by the Moscow Department of health on the usage of AI services in the work of Department of Diagnostic Radiology. Our development of сelcus also takes part in it. The results of the experiment will form the basis for national standards regularizing the usage of AI in clinical medicine.

Despite the presence of a large number of new and specific problems, with the right approach, AI services can already have a positive impact on our lives and health. On the condition that developers and the medical community work in close cooperation, we will hear about real success stories in the near future.