A team of researchers from the Kaunas University, in Lithuania, have developed a deep learning algorithm able to predict the possible onset of Alzheimer’s disease from brain images with greater than 99% accuracy. The method was developed during the analysis of functional MRI images obtained from 138 subjects and achieved better results in terms of accuracy, sensitivity and specificity than traditional methods.
There Research was published in the scientific journal Diagnostics.
Algoritmo predicts Alzheimer’s: here’s what the research says
According to the World Health Organization, Alzheimer’s disease is the most frequent cause of dementia, contributing up to 70% of dementia cases in total. Around 24 million people are affected worldwide and this number is expected to double every 20 years. Due to the aging of society, the disease will become an economic burden on the public health system.
“Medical professionals around the world seek to raise awareness of early Alzheimer’s diagnosis, which gives those affected a better chance to benefit from treatment. This was one of the most important issues when choosing a topic for Modupe Odusami, a PhD student from Nigeria “, he claims Rytis Maskeliūnas, researcher at the Department of Multimedia Engineering, Faculty of Computer Science, Kaunas University of Technology (KTU), Ph.D. of Odusami, supervisor.
One of the possible first signs of Alzheimer’s is short cognitive impairment (MCI), which is the stage between the expected cognitive decline of normal aging and dementia. Based on previous research, functional magnetic resonance imaging (fMRI) can be harnessed to pinpoint brain regions that may be associated with the onset of Alzheimer’s disease, according to Maskeliūnas. The early stages of MCI often have no clear symptoms, but in some cases they can be detected by neuroimaging.
This process is theoretically possible but the manual analysis of fMRI images that attempt to identify changes associated with Alzheimer’s not only requires specific knowledge, but also takes a long time: the application of deep learning and other artificial intelligence methods can accelerate this with a significant margin of time. Finding the characteristics of MCI does not necessarily mean the presence of a disease, as it can also be a symptom of other related diseases, but is more of an indicator and a possible help in orienting towards a doctor’s evaluation.
“Modern signal processing allows you to delegate machine image processing, which can complete it fairly quickly and accurately. Of course, we dare not suggest that a medical professional should ever rely on a one hundred percent algorithm. Think of a machine as a robot that can do the most tedious task of sorting data and researching features.In this scenario, after the computer algorithm has selected the potentially affected cases, the specialist can examine them more closely and in the end everyone benefits as diagnosis and treatment reach the patient much faster “, says Maskeliūnas, who oversaw the team working on the model.
The deep learning-based model was developed as a fruitful collaboration of leading Lithuanian researchers in the field of artificial intelligence, using a modification of the well-known ResNet 18 (residual neural network) to classify functional MRI images obtained from 138 subjects. The images fell into six different categories: from healthy to mild cognitive impairment (MCI) spectrum to Alzheimer’s disease. In total, 51,443 and 27,310 images were selected from the Alzheimer’s Disease Neuroimaging Initiative fMRI dataset for training and validation.
The model was able to effectively find MCI characteristics in the given dataset, achieving the best grading accuracy of 99.99%, 99.95% and 99.95% for early MCI versus AD, late MCI versus AD and MCI versus early MCI, respectively.
“While this was not the first attempt to diagnose the early onset of Alzheimer’s from similar data, our main finding is the accuracy of the algorithm. Obviously, such high numbers are not real performance indicators in real life, but we are working with medical institutions to get more data “, says Maskeliūnas.
According to the scholar, the algorithm could be developed in software, which would analyze the data collected by the groupsvulnerabilu (those over the age of 65, with a history of brain injury, hypertension, etc.) and would notify medical personnel of early-onset Alzheimer’s related abnormalities.
“We have to make the most of the data “, states Maskeliūnas, “That is why our research group focuses on the European principle of open science, so that anyone can use our knowledge and develop it further. I believe that this principle contributes considerably to the progress of society ”.
The lead researcher, whose main area is focused on the application of modern artificial intelligence methods on signal processing and multimodal interfaces, says that the model described above can be integrated into a more complex system, analyzing several parameters, for example , also eye movement tracking monitoring, face reading, voice analysis, etc. Such technology could then be used for self-control and warning to seek professional advice if something causes concern.
“Technologies can make medicine more accessible and cheaper. While they will never (or at least not soon) truly replace the doctor, technologies can encourage the search for timely diagnosis and help.“, Says Maskeliūnas.
Algorithm predicts Alzheimer’s: artificial intelligence is also based on fMRI
Functional magnetic resonance imaging (fMRI) is a non-invasive diagnostic technique for brain disorders. It measures the slightest changes in blood oxygen levels within the brain over time, providing information on the local activity of neurons. Despite its advantages, fMRI has not been widely used in clinical diagnosis. The reason is twofold. First, the changes in fMRI signals are so small that they are overly susceptible to noise, which can nullify the results. Second, fMRI data is complex to analyze. This is where deep learning algorithms come into play.
Network activation map from the output of the second temporal convolution layer mapped to the MNI brain atlas. Credit: doi 10.1117 / 1.JMI.7.5.056001.