Method developed at IC-Unicamp uses neural network to process and analyze brain images taken by magnetic resonance imaging
In Brazil, Alzheimer's Disease (AD) affects 1,2 million people, a number close to that of the population of Campinas (SP). Characterized by the progressive and irreversible loss of cognitive functions, AD is the main cause of dementia among the elderly. Each year, 100 new cases are registered in the country. Although Alzheimer's has no cure, its effects, such as memory loss, mental confusion and impaired sense of direction, can be delayed, as long as the diagnosis is made early. It was with a view to making the identification of AD faster and more accurate that researchers from Unicamp's Computing Institute (IC) designed a technique based on an artificial neural network. Using three-dimensional images generated by magnetic resonance imaging, the method analyzes the morphology of the brain and indicates whether or not the patient has Alzheimer's or whether they present characteristics that indicate the possible emergence of the disease in the future.
The technique was created in the context of computer scientist Guilherme Adriano Folego's master's thesis, under the guidance of professor Anderson Rocha. The work was co-supervised by professor Marina Weiler, from the National Institutes of Health (NIH), in the United States. “Here, we transform pixels into information. In other words, we process and analyze images with the purpose of extracting data of interest. In this case, we seek to identify morphological characteristics of the brain that are compatible with those of the organs affected by Alzheimer's”, explains the professor. Due to cell and neuronal death caused by the disease, the brain loses white and gray matter and takes on a different configuration than the healthy organ.
According to Guilherme Folego, the objective of the technique is to identify structural biomarkers of AD. To do this, he chose to use the convolutional neural network, which allows for deeper machine learning. “In addition, the neural network makes the process fully automatic and faster. While our method provides results in around 15 minutes, in the conventional model, which depends on human intervention, this can take 15 to 20 hours”, he points out.
The technique developed at IC-Unicamp makes the diagnosis in three different classes: healthy person, with AD and in an intermediate stage, that is, those with mild cognitive impairment. “The diagnosis of this intermediate stage is considered critical by Medicine, because it is very difficult to make. If this identification is possible, it is also possible to start early treatment that delays the consequences of Alzheimer's as much as possible”, points out Professor Anderson Rocha.
Another important fact is that the method does not use any information other than the images generated by magnetic resonance imaging. “We don’t know, for example, whether the image is of a man or a woman, how old the patient is or whether there are cases of Alzheimer’s in the family. Our only parameter is brain morphology. The challenge is to understand the subtleties, in order to anticipate the diagnosis, especially in relation to patients in an intermediate stage of the disease”, explains Guilherme Folego.
Professor Anderson Rocha is keen to note that the technique is not intended to replace the work of the neurologist, but rather to help him formulate a more accurate diagnosis of AD. “Quite the contrary, the presence of the doctor is essential, especially because he is the one who will have the final word. We developed an algorithm with the function of learning image patterns and, thus, providing medical professionals with more refined information for decision making. In the tests we carried out, the technique was correct in 75% of cases, which is the same percentage of correct answers as doctors. The sum of these indices plus the contribution of additional information, such as the patient's age, gender and family history, will certainly give greater accuracy to the diagnosis”, he considers.
According to the researchers, intellectual protection over the technique was not requested because the idea is precisely to offer a tool that can be widely used and bring benefits to society. As for the possibility of the method becoming a commercial product, they highlight that before this happens, new tests will need to be carried out. “One possibility is to use magnetic resonance images from patients at the Hospital de Clínicas da Unicamp, whose diagnosis is already known, to evaluate the performance of the tool. Furthermore, we will obviously need a partner willing to transform our technique into a product”, says Professor Anderson Rocha.
The researchers are currently completing an article about the work to be submitted to a high-impact scientific journal. Anderson Rocha notes that the research had important support from Microsoft, which provided its computational structure to run the system. “This is one of the challenges of working with a large volume of images, a procedure that requires very robust computational power”, adds Guilherme Folego.