Using machine learning, researchers at Berlin's Charité Hospital are trying to uncover the mystery of Alzheimer's, depression and MS.

Every 100 seconds a person in Germany falls ill with dementia. Most are affected by Alzheimer's disease. Around 1.7 million Germans currently live with the "people's disease". By the year 2050 the number, according to current projections of population development in an ageing society, will rise to three million. The frightening upward trend is also reflected in other disorders of the central nervous system such as Parkinson's, stroke, epilepsy, depression, anxiety disorders or multiple sclerosis. Today, neurological diseases are already the number one cause of disability in Europe and the number two cause of death, the European Academy of Neurology (EAN) points to the growing threat. It is precisely this mastery that has not been achieved in recent decades, despite a high level of research and promising innovations: Not only is there a lack of understanding of the exact mechanisms, but an early diagnosis of neurodegenerative diseases is also usually not possible - not to mention therapy.

What neurology has not yet succeeded in doing, several research groups at the Charité in Berlin are working on at the same time, focusing on artificial intelligence methods in order to discover the secrets of the human brain and the disorders that occur there. The focus is on neuroimaging. The approach of presenting the central nervous system, especially the brain, as a 2 or 3 dimensional image for better understanding is not new. On the contrary: since the ‘70s, such images have been used by doctors to decide what disease a new patient has. The method of magnetic resonance imaging (MRI) is mainly used. "Depending on the sequence, MRI can be used to display different brain properties," explains Prof. Dr. rer. nat. Kerstin Ritter, junior professor for computational neuroscience and head of the research group "Machine Learning in Clinical Neuroimaging" at the Charité in Berlin. "For example, metabolic properties such as the concentration of metabolites or functional properties such as brain activity in rest phases or during the performance of mental tasks become visible.” However, morphological properties such as tissue loss or inflammation centres are also displayed. These can be signs of neurological diseases.

^{Prof. Dr. Kerstin Ritter © Charité Berlin}

When subtle changes become visible...

But the matter is not always so clear. "While neurological diseases show more or less clear morphological changes, the changes in mental diseases are far more subtle," explains Ritter, because "in some cases they can only be found in the functional area" Accordingly, it is difficult for doctors to detect depressive phases or other mental disorders early on and to diagnose them unambiguously. After all, depression or anxiety disorder have long been a mass phenomenon: According to the World Health Organization, one in three people suffers from such a disease at least once in their lifetime. Nationwide, more than one in four adults in a year meets the criteria for mental illness. This is where machine learning comes in: "they are powerful instruments for detecting even the smallest changes in the brain at an early stage," the expert knows from her own investigations. After studies on depression, follow-up projects are planned in the area of addiction and schizophrenia, according to Ritter. In addition, in another project, the team is trying to filter out mental health indicators through machine learning. It uses artificial intelligence to analyse 15,000 MRI images from the UK Biobank, a globally recognised database for serious and life-threatening diseases.

Decision criteria data

It's a good start. However: "in order to exploit the full potential of machine learning, we need large data sets that map the various diseases in all their complexity and that allow us to learn disease-related deviations," explains the mathematician, appreciating the advantages of the Charité. The large medical centre in the heart of Berlin offers research the outstanding availability of its own high-quality image data, with deep learning methods being particularly successful. These so-called Convolutional Neural Networks (CNN) are deep neural networks capable of hierarchically decomposing data and learning complex relationships. "For example, these networks can be used to detect inflammation sites in multiple sclerosis," adds Ritter, "or to make image-based differential diagnosis between different forms of dementia."

Convolutional Neural Networks are also used in the machine learning research group - one of seven in the Imaging and Neurotechnology research area at the Clinic for Psychiatry and Psychotherapy (Charité). Kerstin Ritter's team has developed a transparent deep learning method that not only enables the diagnosis of neurological diseases based on MRI images, but also explains why the algorithm came to this diagnosis. "We compared different visualisation methods with each other and quantified them in terms of different metrics," reports the Rahel-Hirsch fellow and recipient of the “NARSAD Young Investigator Grant” from the “Brain & Behavior Research Foundation”, "we are currently looking at whether we can identify subgroups from the explanations, and to what extent these explanations with morphological changes, for example hippocampal atrophy in Alzheimer's disease (note: a change in the hippocampus can be an indication of the disease).”

Developed and used for the investigation of Alzheimer's disease, the procedure has now also been used for multiple sclerosis (MS). "Here we have shown that CNN models pre-trained on Alzheimer's data can be very well adapted to data from patients with MS," said Ritter, who would like to pass on her knowledge in the next semester in the compulsory elective module "Artificial Intelligence of Medicine" in the Medicine course at the Charité, “and that she does not only include centres of inflammation in decision-making, but also the localisation of the centres of inflammation and inconspicuous brain areas.” Supported by a research grant from the "German Multiple Sclerosis Society (DMSG)", the next step of the research group will deal with the differential diagnosis of MS and the prediction of disease activity in individual patients.

MRI 4.0

When dealing with the autoimmune neurological disease MS, in which the immune system attacks the brain, Kerstin Ritter's team takes advantage of its integration into the Charité. The data on both the previous and the planned research projects originate from the company's own resources: From the "Translational Neuroimaging Laboratory" of the Clinical Neuroimmunology research group. It explores new and improved imaging techniques to improve the follow-up of MS and thus the treatment of individual patients. "An important disease element of multiple sclerosis are lesions (note: injuries, disorders) in the brain, which we can detect by magnetic resonance imaging," says Prof. Dr. med. Alexander U. Brandt, head of the research group, citing an example that shows the potential of machine learning, "but in individual cases patients can have hundreds of these lesions that spread three-dimensionally across the brain. We know that measuring the number, location and size of these lesions can give us important information about the diagnosis, prognosis and course of the disease. But to measure these lesions manually is simply no longer possible and certainly not practicable in the clinic. Using artificial intelligence, we can automate this process and integrate the valuable results of this otherwise very time-consuming work into everyday clinical practice."