The Alzheimer’s has no cure and is the leading cause of dementia in the world. This disease affects more than 50 million people in the world. Although those who suffer from it cannot overcome it, the early detection is considered key in order to develop effective treatments. Mild cognitive impairment is a phase that precedes the disease, but not all people who suffer from it end up developing Alzheimer’s.
Research led by scientists from the Open University of Catalonia (UOC) and published in the journal IEEE Journal of Biomedical and Health InformaticsHas got distinguish with great precision those in which the deterioration is stable and whonevertheless, will progress to disease. All this is achieved by means of an artificial intelligence that uses magnetic resonance image recognition, which is more efficient than the rest of the methods used today.
“These patients can progress and worsen or stay in the same state over time. Therefore, it is important to distinguish between progressive and stable cognitive impairment in order to prevent the rapid progression of the disease”, explains Mona Ashtari-Majlan, a UOC researcher in the AI for Human Wellbeing group.
To achieve this, the researchers used a method called multi-stream convolutional neural network, an artificial intelligence and deep learning technique very useful for image recognition and classification.
“We first compared MRIs of Alzheimer’s disease patients and healthy people to find different benchmarks,” explains Ashtari-Majlan. After to train the system, they adjusted it with resonance images of people who had already been diagnosed with stable or progressive cognitive impairment and in which the differences are much smaller. Altogether, almost 700 images from public databases were used.
In addition, the proposed method could solve the problem of small sample size, since the number of MRIs for cases of mild cognitive impairment is much lower than for Alzheimer’s.
The new method allows distinguish and classify the two forms of mild cognitive impairment with an accuracy close to 85%. “The evaluation criteria show that our method outperforms existing ones”, confirms the researcher, including more conventional methods or others based on deep learning. (YO)