AI fairy next to the sick bed

For medical imaging, IT is not entirely new. We can even confirm that there are algorithms in digital imagery because, without them, it would be impossible to utilize the actual signals provided by the device, MRI and others, completely avoiding our meaning and creating images of the human body. However, in the last four to five years, with the advent of deep learning algorithms, a new step has been taken in the computer processing of digital medical images, successfully using huge amounts of data that no human brain is able to fully comprehend.

Recent examples have left a mark on my mind. In dermatology, deep learning software, previously trained on more than 1 million annotated generic images, adapted 130,000 dermatological abnormalities to how to automatically differentiate melanoma from benign lesions with the skill of a dermatologist. In radiology, the French company Therapixel submitted its software to the proposed 640,000 mammograms during a global competition and won by identifying the best suspicious images from all its competitors. This program is measured today by specialist radiologists. In pulmonology, software developed by Google, trained on thousands of scanners and tested on over 1,000 new images, is capable of detecting the presence of suspicious nodules in lung volume scanners and identifying them as a professional.

Finally, ophthalmology, other learning software, trained on more than 130,000 images of the retina, reliably detects diabetic retinopathies with the reliability of an ophthalmologist. This is in the case of IDx-DR, approved by the FDA, American Drug Agency in 2018, so that the disease can be diagnosed automatically without looking at pictures by a specialist. Since then, more than a dozen AI-assisted imaging systems have received this label.

A digital double for each patient

While these results are spectacular, they are currently limited to specific, relatively narrow tasks, for which a huge database filled with experts is available. But these are not database forces, their acquisition and their analysis by experts in particular is very costly. Furthermore, in case of doubt, deep learning algorithms are characterized by opaque millions of parameters for interpretation. They are also susceptible to counter-attacks. In 2019, a survey revealed that added to the images, small disturbances completely invisible to the naked eye but very accurately calculated disturb the results. Under this condition, a skin lesion initially classified as benign with 100% confidence becomes malignant with the same level of confidence, where the difference disappears with the naked eye. The reverse is also possible, and more serious, since a hacked database can lead to the absence of cancer treatment.

How to make AI algorithm more powerful? One way is to integrate them with the patient’s digital models, the other using common sense in our anatomy and physiology, but confines itself to a small number of more easily explained parameters. More specifically, algorithms of living organisms and mathematical, physical and biological models now make it possible to create, from medical images and other related data, a digital and personalized representation of the patient, a “personalized digital”, “e-patient”. , Or “Digital Twin”. The parameters and models of these replicas can then be used by digital medicine algorithms designed for diagnosis, prognosis and therapy.

Helps to enrich the database by simulating digital patient biophysical models, For example, in rare diseases or cases which are slightly or poorly presented among the listed patients. Carried out intensive simulation Descending The mass of different information is useful for speeding up scientific calculations in new cases. The statistical approach to machine learning therefore complements the more determinative approach.

Finally, digital patient theoretical frameworks and modern methods of big data processing favor data integration that can be described as overall on the patient, including not only their physiological and functional picture, but also very complete biological information. (Genetics, Metabolism 7), depending on the behavioral and even the environmental lifestyle (we then refer to the “exposome”).

Analyzing all these heterogeneous and large-scale data is a scientific challenge, but research is very active on these issues, especially within the framework of 3IAs (Interdepartmental Artificial Intelligence Institutions) deployed in France. The aim is to advance the understanding of diseases and to better manage them, especially in the detection and treatment of neurodegenerative diseases, including lung and breast cancer.

Finally, the field of application of artificial intelligence in medical imaging will be extended to all medical disciplines (radiology, cardiology, oncology, neurology, radiotherapy, surgery, etc.). As a result, doctors and all health professionals, especially pharmacists, medical professionals, etc., must be trained in AI to better understand its usefulness and its limitations. And some are already developing university degrees dedicated to 3IA (in the case of C ডিte d’Azur in Nice-Sophia Antipolis).

Gather information

Extensive information exchange is also required. In the United Kingdom, the UK-Biobank database monitors the scientific community over time with over 100,000 volunteer participants gradually acquiring all the information (images, genetic analysis and environmental parameters). In the United States, the ADNI database offers, for a large group of Alzheimer’s patients, anonymous images of their brains, as well as clinical and genetic data.

France launches more: Health Data Hub (HDH), AP-HP Health Data Warehouse (EDS), France Life Imaging, Dream France IA, etc. To protect and make them available to researchers with participants’ agreement. Projects necessary to anticipate progress in research, perhaps significant, in research, and for that we must ensure that regulations, while adequately protecting patients, do not penalize French and European researchers compared to the rest of xfworld.

One way to do this is to develop federated learning. The data remains at the centers that created them, and it is the learning algorithms that work on each of the centers’ computers and that share local learning between them to provide an overall result. Leading such a federated learning project for nuclear medicine imaging, our 3IAs have partnered with a number of centers to fight cancer, including the Antoine-Lacassagne Center in Nice, the Curie Institute in Paris and the Henri-Becquerel Center in Rouen.

Finally, artificial intelligence, medical imaging and digital patient “4 Ps” create a set of drug-serving IT tools: more personalized, precise, predictive and preventive. These digital tools will save the therapist a lot of time to take better care of their patient. They will assist general practitioners by proposing to refer patients to specialists if necessary. These new tools are intended to support the doctor, not to replace him.

Leave a Comment