Data vs. Dementia
Christian Wachinger is using machine learning to predict early stages of neurodegenerative diseases based on large-scale neuroimaging data.
Some 47 million people now suffer from dementia, and forecasts suggest that this number will rise to over 140 million by the year 2050. These figures underline the need to identify early indicators of neurodegenerative disorders, as this would increase the chances of slowing the progression of these devastating conditions. Computer specialists believe that artificial intelligence (AI) could make a crucial contribution to the early diagnosis of brain diseases such as Alzheimer’s. “Machine learning can help doctors to detect the earliest signs of Alzheimer’s in brain scans obtained by magnetic resonance tomography,” says LMU neuroanatomist Christian Wachinger. “Algorithms can learn to recognize patterns and subtle anatomical changes in MRT scans that are very difficult to discern with the naked eye.” So AI has the potential to extend the capacity of medical diagnostics and, in so doing, it could transform the whole field.
However, developing reliable algorithms requires access to enormous amounts of training data. This is how today’s Go programs learned to play the game better than the world’s recognized masters. “Programs like AlphaGo Zero, developed by DeepMind, were able to improve so fast only because they were repeatedly tested against one another, and could continually enhance their strategies by being steadily confronted with novel situations and configurations,” says Wachinger, who heads the AI Laboratory for Medical Imaging in LMU’s Department of Child Psychiatry. “In the clinical context, we need thousands of datasets for every situation obtained from real patients in order to train the systems optimally.”
In practice, the amounts of data needed exceed those collected in all but the largest medical centers. This has led to the formation of international research collaborations, whose members obtain access to specialized databases, such as that assembled by the Alzheimer’s Disease Neuroimaging Initiative (ADNI). ADNI includes data from approximately 2000 individuals who have their brains regularly scanned by MRI or PET (Positron Emission Tomography) and whose genomes have been sequenced. Most of the subjects exhibit memory defects, but healthy persons also contribute data to ADNI.
“Medical databases of this sort are a great boon for us,” says Wachinger, who has trained his algorithms on over 6000 brain scans. All relevant data – age, detailed medical findings and genetic markers – are encoded as points in what is known as a high-dimensional vector space. During training, algorithms are confronted with similarly encoded data derived from MRI images. The algorithms are organized into hierarchical neural networks. Each level recognizes specific features in the digitized images – simple geometric shapes, patterns, curvatures. Each node in a network is connected to other nodes and other levels in the stack. The weightings assigned to nodes are constantly adjusted. This essentially reflects how specific connections between neurons in the brain are reinforced by experience during learning, and explains why these algorithms can progressively improve the quality of their predictions.
Wachinger’s algorithms can distinguish healthy individuals from patients with diagnoses ranging from mild cognitive deficits to vascular dementia. Moreover, they can assess the likelihood that a patient will develop overt signs of Alzheimer’s disease within 6 or 12 months.
The primary determinant of the accuracy of such prognoses is the choice of the biomarkers employed, i.e. structural features that are correlated with a particular clinical outcome. These include the thickness of the cerebral cortex and the volume of subcortical structures, as well as the detailed geometry of these structures. The use of the latter represents a new departure. Their spatial geometry, or shape information, can be captured in terms of oscillation patterns, which give rise to different resonance frequencies. Changes in the form of these structures can serve as an early indicator of asymmetrical loss of neurons. Hence, the ability to detect such alterations is useful for the early diagnosis of dementias, and enhances the precision and the speed with which such predictions can be made.
This new approach to the analysis of the neuroanatomy of particular brain regions has enabled Wachinger to show that the hippocampus (which is involved in memory formation and the triggering of emotions) and the amygdala undergo asymmetric changes in Alzheimer patients. “As the degree of dementia increases, these asymmetries become more pronounced.” Conversely, in healthy controls, they are barely discernible. “The asymmetries observed in these structures could be an early-stage biomarker for Alzheimer’s,” Wachinger concludes.
Dividing the brain into clearly defined anatomical regions on the basis of an MRT scan is no easy task. “The scan is essentially made up of more or less gray dots.” Wachinger employs machine learning to digitally define the various brain regions and cleanly separate them one from another. “The 10 million or so pixels that make up an MRI brain scan are converted into about 100 image-derived measures.” The compression procedure is the core of the technology, and it is based on a highly detailed knowledge of both the medical and the technical aspects of bioimaging techniques. “As a computer specialist, I am primarily interested in the methodological aspects of the process,” says Wachinger. “Nevertheless, it’s marvelous to be able to contribute to advances in medicine and to help people.”
Identifying brain structures is based on a neural network model that builds on existing software such as FreeSurfer, which was developed at Harvard (Wachinger did his postdoc at MIT and Harvard Medical School). His team has now written a program which was released early this year. Dubbed QuickNAT, it divides the brain into anatomical segments and estimates their volumes – all within 15 seconds. So the clinician receives the algorithmic diagnosis while the patient is still in the scanner. “The procedure can be incorporated into the everyday routine relatively quickly. As in the case of a blood test, this would provide the physician with quantitative measures of the neuroanatomy of each patient’s brain,” says Wachinger. How this information is used is a matter for the clinician concerned. In order to make the procedure widely available, Wachinger has created a web-based service for doctors, which allows them to upload MRI scans for evaluation.
Despite these advances, Wachinger warns that it is too early to take such algorithmically based conclusions at face value. “The methods we use can only provide probability values, together an estimate of the level of uncertainty associated with the procedure. “Therapeutic decisions based on such data remain in the hands of the consulting physicians,” Wachinger emphasizes once again. Why Alzheimer’s patients exhibit such asymmetries in the sizes of the amygdala, the putamen or the hippocampus is still unknown.
“More like a rock concert …”
To identify possible grounds for these asymmetries, Wachinger’s team recently made use of a type of genetic data known as ‘single-nucleotide polymorphisms’ (SNPs). The term refers to population-level variation at single, defined sites in human genomes, and SNPs are now used as training data for algorithms that sift through medical datasets. Some SNPs affect the level of risk for certain neurodegenerative diseases, and Wachinger’s group has now linked two novel SNPs to the anatomical asymmetries mentioned above. Their findings may well prompt others to search for the functional basis for these links.
Given the euphoria surrounding the subject, Wachinger is at pains to avoid raising false hopes. “In light of the flood of publications, skepticism is advisable with regard to the quality of AI-based methods.” He views the current boom – with people piling in on all sides – with some amusement, noting that the annual conference on Neural Information Processing Systems in Montreal in 2018 was booked out in just over 10 minutes. “The event is more like a rock concert than a science conference,” he adds. “The whole field is exploding, which makes it difficult to distinguish the relevant from the incidental.”
Nevertheless, AI in medicine has tremendous potential. Algorithms can be trained to detect diverse diseases, such as autism, diabetes or psychiatric conditions such as depression. “All we need are large, informative datasets and good biomarkers – in the case of diabetes, imaging data for the liver and the kidneys.” He is already in contact both the UK Biobank, which is collecting whole-body MRT scans from 100,000 individuals, and the National Cohort, a long-term health study now underway in Germany. Such datasets are grist for his mills – sorry, algorithms. Hubert Filser
Dr. Christian Wachinger heads the Laboratory for Artificial Intelligence in Medical Imaging in the Department of Child Psychiatry at LMU, which is supported by funding from by the Bavarian State Ministry of Education, Science and the Arts in the framework of the Center Digitisation.Bavaria (ZD.B). Wachinger (b. 1982) studied Informatics and obtained his PhD at the Technical University of Munich. He subsequently did a postdoc as a member of the Medical Vision Group in the Computer Science and Artificial Intelligence Lab at the Massachusetts Institute of Technology (MIT) in Cambridge, and in the Laboratory for Computational Neuroimaging at Harvard Medical School in Boston.