Artificial intelligence, assistant to the radiologist
Artificial intelligence has been at the heart of concerns in recent years, particularly in the field of medical imaging. It raises many questions and raises debates about its ability to replace radiologists, but what is it really?
Under the name artificial intelligence (AI) we combine all the theories and techniques implemented in order to create machines capable of reacting and simulating intelligence. This practice allows a computer system to solve complex problems by integrating logic.
The concept of artificial intelligence was born in the 1950s, thanks to the mathematician Alan Turing, who raised the question of bringing a form of intelligence to machines. In recent years, artificial intelligence has grown rapidly due to advances in research. Its scope has been considerably diversified, ranging from autonomous cars to smartphones to robotics. Now many areas are undergoing the upheavals associated with this new technology. Health, especially medical imaging, is no exception and now must deal with artificial intelligence.
The arrival of AI in medical imaging is the result of three major developments:
- The digitization of medical images allowing them to be setup.
- The development of algorithms allowing the use of data captured in natural language.
- The emergence of deep learning allowing, from massive radiological data, to develop algorithms for automatic analysis of medical images.
Those solutions make possible today automatic lesions detection and can be integrated into large cancer screening programs (lung, breast, prostate...)
In order to fully understand what the term artificial intelligence means; it is necessary to dwell on how to produce it. Algorithms are at the heart of the creation of artificial intelligence. They allow systems to enter a learning process, in order to accomplish a task or a complex set of tasks, this is called machine learning. There are a significant number of algorithms that can be used in this framework. We can group them into three large families:
- Supervised learning algorithms, in which input data, also known as drive data, are annotated and have a known result. A model is created through this training process in which it is necessary to make predictions as well as corrections.
- Unsupervised learning algorithms, for which input data are not annotated and have no known results. The system creates its own model by identifying the structures in the input data.
- Semi-supervised (hybrid) learning algorithms, where input data is a mixture of annotated and raw data.
Just over two years ago, artificial intelligence made a notable entrance for RSNA 2017. Many radiologists feared the death of their profession in favor of software that has become so much, if not more, and sometimes more efficient than they are. In 2020 some of the fears and uncertainties associated with this intelligent software are still present but mitigated. Other fears have emerged, for example about the performance, use and clinical validation of some of these artificial intelligence algorithms.
Radiologists today face a twofold problem:
- The workload is constantly increasing, challenging their ability to provide patients with quality interpretation and reasonably required attention.
- The increasing complexity of the available data as well as their volume.
The finding is unmistakable. Radiologists face great difficulties in carrying out their daily tasks and are finding it increasingly difficult to meet the demands of other clinicians and patients. It is in this challenge imposed on the radiologists that artificial intelligence takes on its full meaning. This challenge is to provide the patient's benefits with greater quality while promoting a return to a more human relationship. The context requires these clinicians to work faster, while performing more complex interpretations at a lower cost. The increasing use of radiology in all clinical disciplines (therapy, screening and intervention) pushes the radiologist to provide more complete information (radiomic and radio-genomic analysis, activity curves, fusion data...).
To meet these challenges, the radiologist needs short-term help. Artificial intelligence can be a tremendous boon if used appropriately, being able to increase the effectiveness of medical imaging professionals tenfold. Artificial intelligence will improve the process of programming and conducting reviews, and data analysis assistance will be equally useful. Automating repetitive tasks will free up more time, which can be spent on more sophisticated analysis. In a study carried out by the French Radiological Society (SFR) in 2017, 82% of medical imaging professionals already said that artificial intelligence could improve the relevance of clinical decisions and 80% of them said it could optimize their productivity. Artificial intelligence therefore remains a tremendous opportunity for physicians to free up time for improved follow-up and patient relationships.
Marketplace or integration?
There are now two different industrial approaches to the provision of AI for health professionals.
The first approach is that of the marketplace, which consists of providing, on a platform, several clinical applications classified by criteria of specialties, according to modalities or even by anatomical regions. These marketplaces work in a similar way to what can be seen on smartphone app platforms at Apple or Google. In our view, this approach raises some problems. The main question is the choice to be made by the clinician. When he is faced with several clinical applications that perform the same function. On what criteria should you base your choice? What is the best application, which will be the most effective and clinically relevant to complete an examination? How will it fit into its workflow?
We consider that the role of the radiologist or clinician is not to make these technical choices. It is up to us, as an industrialist, to make available the most efficient and relevant tools when they need them.
To do this, we are collaborating with leading players in the field of artificial intelligence. Our approach is to conduct pre-selection based on clinical evaluations conducted with radiologists and clinician partners. We then define with them which application is the most efficient and best meets clinical needs.
We have also chosen to integrate these artificial intelligence algorithms into a clinical, fast and intuitive workflow that adapts to the needs of the radiologist. Indeed, during his interpretation work, the radiologist finds himself in a very structured diagnostic process, step by step. Our workflow is modeled on this analysis process, allowing the radiologist to have the right image displayed in the right way and at the right time, with the most relevant information associated with the right tools. Our work and know-how are based on this ability to integrate artificial intelligence into the radiologist's clinical workflow, to make it relevant to the healthcare professional in the service of the patient's care path.