Lithuanian researchers from Kaunas University of Technology (KTU) improved an algorithm that detects Alzheimer’s disease from magnetic resonance imaging (MRI) images. The new model achieved over 98 per cent accuracy on a test dataset in detecting the neurodegenerative disease by improving a neural network model.
Alzheimer’s disease (AD) is the seventh in the U.S. and one of the leading causes of death in the world. Patients with AD often experience memory loss and cognitive decline due to the impairment and death of nerve cells in the brain. Usually, to diagnose this disease a psychiatric evaluation has to be performed, memory and problem-solving skills must be tested, or various brain scans, including MRI, have to be performed. Detecting an early stage of AD is an especially difficult task.
To facilitate the process of diagnosing AD, the KTU researchers developed a deep learning method to detect early signs of AD from MRI images. The model followed the original idea of their previous study but used a modified algorithm and a broader network to achieve more adaptable results.
The latest studies have shown that pre-trained convolutional neural networks (CNN) can accurately diagnose cognitive disease from brain magnetic resonance images. Rytis Maskeliūnas, a researcher at the KTU Department of Multimedia Engineering, says that the previous study was based on the modification of the ResNet18 network, so this time we have investigated a modified variant of the DensNet201 network, which has better parameter optimisation.
“The previous study had less flexibility: When we train CNN with specific data sets, such as MRI images, we cannot accurately account for variations in data, the algorithm expects to get the same style and format of images all the time. In reality, however, the patient may visit different hospitals, that will have different equipment for taking the MRI, it might be adjusted or configured differently, a person may lie on the scanner bed differently, so the photos will be ever come out a little different,” explains Maskeliūnas.
A collection consisting of images of brain scans from 125 subjects from The Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset was used for the study. Images were analysed in terms of Alzheimer’s disease, mild cognitive impairment, and dementia. The data set utilised in the investigation is open and constantly updated with the latest images of AD patients, so the results of the study are up to date and relevant.
“Using the ever-increasing ADNI dataset, the algorithm is getting ready to recognise the symptoms of the disease in various images and becomes less sensitive to a specific data source. It’s not a revolution, but certainly an evolution,” says the KTU researcher.
Apart from the use of an additional network and ADNI dataset, the study differs from previous research by using a different weight mechanism and employing a modified gradient class activation map. It is a step forward towards practical application because the model will soon be able to mark affected areas of the brain.
“We might soon use this research in medical fields. Our goal is to create a model that spots the symptoms of AD in the brain and marks the affected area on the computer screen, helping the medical professional to examine the image. So, by including new parameters and more broad data sets we are improving this model,” says Maskeliūnas.
KTU researcher notes that in the future more variables could be added to the study to speed up the process of diagnosing.
“In the future, we plan to use biological markers and other brain scanning methods for even greater diagnosing efficiency and better adaptability,” says the researcher.
For the development of the algorithm Maskeliūnas and his PhD student, Modupe Odusami were awarded the Discovery of the Year title in the Lithuanian national awards Lithuanian Honour. This study is one of the several related to the health tech sector and conducted by Maskeliūnas and his team.
Reference: Odusami M., Maskeliūnas R., Damaševičius R. An Intelligent System for Early Recognition of Alzheimer’s Disease Using Neuroimaging. Sensors. 2022, 22(3). https://doi.org/10.3390/s22030740