Artificial Intelligence in medical imaging: Who? What? When?

It is without doubt that the most promising area of innovation in healthcare is, and will remain, the application of artificial intelligence (AI). AI has the potential to improve clinical outcomes and raise further the value of medical data. Indeed, AI is opening new doors at every step of health management, screening the populations, epidemiology, diagnostics, follow-up, treatment planning, and is also widely used in the pharmaceutical industry. And just as radiologists many years ago were the first to tackle the digital revolution, they will also undoubtedly now become the avant-garde of AI in the medical arena.

With an irreversible increase in the amount of data acquired every year feeding all AI applications, the medical imaging market has largely benefitted from this trend in the last 10 years, reaching US$30 billion in 2019, according to Yole Développement (Yole) Artificial Intelligence for Medical Imaging 2020 report. Each second, about 360 medical imaging examinations are being acquired and analyzed worldwide, which equates to more than 2 billion medical imaging exams acquired in 2019. AI in general, and deep learning in particular, can help radiologists extract as much information as possible from this humongous pool of data.

Killer applications for the adoption of AI in medical imaging

Deep learning is a type of AI technology based on artificial neural networks able to detect automatically what it has learnt. This technology was initially implemented for recognition models in images at the beginning of the last decade and have shown extraordinary results since. Medical imaging is a perfect case study for the adoption of AI in the medical industry.

Deep learning models increase in complexity and value in line with the requirement of being more and more precise:

 The higher the added value for both the hospital workflow and patient outcome becomes, the higher the cost of the algorithm. Today, most AI models available in the medical imaging market are screening models aiming at detecting abnormalities or segmenting lesions, but more and more models are developed for automated diagnostics. This will be the biggest trend in the next 5 years. Later on, AI will also be used as a tool to predict treatment outcome and give treatment advice based on medical images…

Deep learning models empowered by AI can also be segmented in terms of imaging modalities. Indeed, AI is used on different types of images: from Magnetic Resonance Imaging (MRI) to Computed Tomography (CT) scanning or X-rays and Ultrasound imaging. However, not every type of modality requires the same algorithm. In fact, modalities can be organized into two types of procedures: quality procedures, which include MRI and CT scans, and fast imaging procedures, which include ultrasound and X-rays.

MRI and CT scans are intensive procedures able to acquire high quality images. With the addition of annotations on the images, the model can reach very high accuracy in classifying pathologies or segmenting objects. Furthermore, the execution speed of the model does not need to be very fast, as the imaging procedure is usually long. On the other hand, models trained on ultrasound images need very fast execution to be able to process real-time images. Those models are then used to detect abnormalities faster and to prioritize cases, implying an important productivity gain.

A software business poised to grow rapidly

Thanks to these powerful tools that deep learning is bringing to the medical imaging field, the AI market in this segment is expected to grow rapidly. Original equipment manufacturers (OEMs) Siemens Healthineers, Philips, and GE Healthcare saw this opportunity early and are already introducing AI tools into their product suites. This trend will gain strength with the introduction of diagnostic algorithms into the market from 2021 onwards. The large productivity gain linked to the use of such models is invaluable, enabling hospitals to more effectively diagnose and treat patients.

Yole’s analysts estimate that the total market in 2025 for software generated revenues through the sale of AI tools will reach US$2.9 billion with a Compound Annual Growth Rate for 2019 to 2025 (CAGR2019-2025) of 36%, shared between the main applications: improved image capture, noise reduction, image reconstruction, screening, diagnostic and treatment planning.

The AI for medical imaging market ecosystem is made up of a myriad of software start-ups, mostly various spin-offs from universities. It highlights the fact that this market is still not stable, driven by historical OEMs in medical research and hardware and Cloud companies that clearly see a nice opportunity there with the need for high performance computing to develop these types of algorithms.

Each software company is highly specific, providing AI algorithms for dedicated applications and modalities. They are able to sell their products directly to hospitals or through two different types of players: marketplaces such as TeraRecon, which are integrating these algorithms into the radiologist’s workflow, or through OEMs such as Siemens, which are optimizing these models in their medical imaging machines.

In fact, OEM machine manufacturers and marketplaces, which are software companies, are offering the same product. Which type of player will be in the strongest position in the coming years? It is still too early to give a definite answer, but all these players are gaining momentum with more and more products coming out and more and more market to catch.

Data is the fuel of AI algorithms but medical data and the medical field in general come with constraints.

Compared to other markets, such as consumer or automotive, the medical field is very restrictive for the development and deployment of AI algorithms. The constraints are either related to the medical data or to regulations and ethical concerns.

Algorithm producer companies find themselves confronted with many limitations when bringing their products to the market. The medical field is, of course, closely monitored in order to avoid the introduction of products that are potentially harmful to patients. One of the key steps in this process is the validation of national administrations such as the FDA and the European Administration (for the CE mark). The process to get those approvals is time consuming and can be hard to handle which represents huge obstacles to overcome especially for smaller companies and software startups.

Before 2018, very few algorithms were approved due to regulations not covering deep learning possibilities. The FDA created in 2018 a specialized team dealing only with deep learning algorithm clearance and since then we have seen an exponential growth in the number of cleared algorithms.

Ethics is also surely one of the most difficult limitations to overcome for deep learning companies. Most people are not ready to trust a computer to deal with their health problems. Furthermore, deep learning is not an easy technology to understand and explain. This complexity strengthens the current reluctance for AI by both patients and caregivers. Radiologists foresaw AI development with trepidation a few years ago. Some possibilities to aid acceptance of AI are a better explanation of the technology, with demonstrations, but also the training of radiologists to assist in deep learning model training, to let them “teach” the models. Development of models that apply only to simple cases, like automating recurring tasks, will probably help artificial intelligence to enter the hospital doors without directly being used to alone determine a patient’s diagnosis. 

Another aspect of the medical data issue is access to this data. Annotated images are the fuel for AI, and access to them is often restricted, as well as other interesting data concerning the patients’ age, gender, size, prior medical history, etc. Some players base their business model solely on data analysis, mainly for advertising use (companies like Google). It seems obvious medical data represents big opportunities for those who can process it efficiently. What limits those companies are the current regulations concerning data privacy leading to the encryption and anonymization of medical data.

In November 2019, a scandal broke out after the revelation of a partnership between Google and Ascension, one of the US’ leading non-profit health systems, allowing google to have access to the un-anonymized medical data of millions of Americans. Having access to the names of patients allows Google to correlate patient outcome with medical history and the evolution of the patient’s physical condition. All this data is used to train the models and allow the execution of diagnostic tools and treatment planning. The interesting point here is the fact that neither patients nor doctors are aware of this data transfer even though this partnership is completely legal. HIPAA allows private actors to share data without informing patients if the information is used to help the entity carry out its health mission.

A cloud-based infrastructure today, a future OEM business tomorrow?

Training needs so much computing resource that it is done in the Cloud by these software startups that want to sell the trained model, called an inference model. This is executed on servers in the hospital for privacy concerns. IT infrastructure directly in hospitals is more secure than cloud computing. The medical data does not leave the hospitals and therefore is protected. The setup of such environments is difficult and time consuming. Nevertheless, it represents the opportunity for hospitals to have full control of their data.

In the future, this technology will probably be capture by OEMs that will push the analysis inside their own machine and directly give the diagnosis to radiologists. This implies two things: inference, even if it less demanding in computing resources than training, will need powerful hardware to enable this technology to be embedded. The efficiency of the technology will have to be proven and qualified by organizations such as FDA to be allowed. However, that opens a new type of business too through software update licensing, which is very valuable. Something we see in automotive, with Tesla for example.

Future of the radiologist profession and impact of AI in hospital workflow

When we first started our study on AI for medical imaging, we had this common idea that AI would sign the end of the radiology profession. Physicians being replaced by robots is a common mistaken thinking of where the future of medicine lies. AI models are transforming the radiologist’s profession, that is true, but for a greater good.  Thanks to AI, the radiologists’ productivity will increase drastically due to an improvement of their workflow through prioritization of critical tasks and the removal of repetitive tasks, which will no longer be performed by the radiologists. This will allow them to focus more on diagnostics and patients. Moreover, the real power today is in the hand of the radiologists as they are the only ones able to make proper annotations to the medical imaging data used for the training of the algorithms and, as we said earlier, data is the fuel for AI. Without data, there is no AI!

In the coming years AI will also have a tremendous impact at the hospital level. Artificial intelligence, and especially deep learning, allows more in-depth analysis as well as autonomous screening in the medical imaging field. Medical images represent a large majority of the global medical data, and due to the large amount of data available, it is possible to develop deep learning algorithms without being limited by the amount of data, which is often the case for AI in many other fields. This new technology will start the transition from volume to value based care. Deep learning tools can look at a large range of data to extract consistent information directly usable by radiologists or other medical professionals. The summarization and extraction features are by far the most important features of deep learning and enable hospitals to decrease time spent on data management. These tools represent opportunities for hospitals to better manage their costs by increasing their productivity. They represent a huge improvement for patients, who will receive faster and more accurate diagnosis and who will be prioritized when any abnormality requires rapid intervention. This will also tend to reduce costs for the patient when accessing healthcare.

About the author

As a Technology & Market Analyst, Medical & Industrial Imaging, Marjorie Villien, PhD., is member of the Photonics & Sensing activities group at Yole Développement (Yole).
Marjorie contributes regularly to the development of imaging projects with a dedicated collection of market & technology reports as well as custom consulting services in the medical and industrial fields. She regularly meets with leading imaging companies to identify and understand technology issues, analyze market evolution and ensure the smart combination of technical innovation and industrial application.
After spending two years at Harvard and prior to her position at Yole, Marjorie served as a research scientist at INSERM and developed dedicated medical imaging expertise for the diagnosis and follow-up treatment of Alzheimer’s disease, stroke and brain cancers.
She presents to numerous international conferences throughout the year and has authored or co-authored 12 papers and 1 patent.
Marjorie Villien graduated from Grenoble INP (France) and holds a PhD. in physics & medical imaging.

Related report

Artificial Intelligence for Medical Imaging 2020

With the emergence of AI in imaging, the medical industry and the radiology profession have begun to dramatically change.

Source: http://www.yole.fr