Is it possible to use deep learning to teach computers to diagnose brain diseases as well as medical doctors?
Imaging technology has revolutionized healthcare and is widely used for diagnosis before treatment or surgery. Despite these advances, routine clinical MRI data interpretation is mostly performed by medical experts. Is it possible to use deep learning to teach computers to diagnose brain diseases as well as or in some aspect even better than medical doctors?
Deep learning is about using powerful computers with embedded artificial intelligence to resemble the human brain’s way of interpreting new information and draw conclusions in relation to what is already known. The difference is that computers, amongst other things, are able to analyse much larger amounts of data, which can be used to find better methods for solving difficult mathematical and technical problems.
“Using a large amount of brain image data, deep learning methods can be used to find characteristic features related to some diseases, and provide powerful diagnostic tools to medical doctors”, says Irene Gu, Professor in the signal processing group at Chalmers.
So far, only preliminary research work on deep learning is reported in the medical area. In computer vision, deep learning has reached or even surpassed human performance when it comes to face recognition.
Recently, Irene Gu has started a research initiative on brain image analytics using deep learning methods in close collaboration with medical doctors at Sahlgrenska University Hospital and several students. The question is: Would it be possible for artificial intelligence technology to diagnose Alzheimers’ disease, or to find brain tumors’ grading, by only using a large amount of brain image data?
“We have obtained some initial promising results. Our ambition is to reach the performance of medical experts and yet in much simpler ways”, says Irene Gu.
Detection of Alzheimer’s disease
Alzheimer’s disease is a chronic neuro-degenerative disease currently incurable, its cause is not yet completely understood. According to WHO’s statistics in 2015, roughly 30 million people in the world suffer from Alzheimer’s. The symptoms consist of disorientation, language difficulties, memory loss, mood swings and many more. Early diagnosis and treatment can potentially slow down the development of the disease.
Brain scans by magnetic resonance imaging, MRI, is a commonly used diagnostic method for detecting Alzheimer’s disease. This is often used in combination with other diagnostic methods involving a set of clinical exams, by observing the progression of dementia symptoms.
“In this project, two dedicated deep learning methods, simple yet effective, have been developed for detection of Alzheimer’s disease. One method is based on 3D convolutional networks, another on 3D multiscale residual networks. We use a large amount of brain MRI scans to learn our computers the features of Alzheimer’s disease, and subsequently to detect Alzheimer’s patients from unseen scans”, Irene Gu explains.
The study involved 340 subjects and about 1200 MR images, obtained from a public available dataset, Alzheimer’s Disease Neuroimaging Initiative (ADNI).
“The proposed schemes have yielded high accuracies. For example, one method has reached an accuracy of 98,74 % on previously unseen MRI scans, and 90,11 % from MRI scans of unseen patients in the study. This almost reaches the highest state-of-the-art research results”, Irene Gu says. “This indicates that the method that we have developed is useful in this type of studies.”
One of the projects was conducted by Mahmood Nazari and Karl Bäckström as a master’s thesis project.
A paper submitted on this work has recently been accepted by IEEE International Symposium on Biomedical imaging (ISBI) 2018. Another MSc project is still ongoing.
Brain tumor grading
Encouraged by the good deep learning results using MR images, Irene Gu has started another project based on similar technology, performed by Karl Bäckström in 2017.
“Thanks to the interest in computer-assisted brain tumor diagnostics shown by medical doctors at Sahlgrenska, and seed funding from the department of Electrical Engineering at Chalmers, we could perform a study on brain tumor (glioma) grading using deep learning”, says Irene Gu.
A glioma is a type of tumor that starts in the glial cells of the brain or the spine. Gliomas comprise about 30 percent of all brain tumors and central nervous system tumors. About 80 percent of all malignant brain tumors are gliomas.
The broad international collaboration networks, which the medical doctors are engaged in, have provided the researchers with brain tumor datasets from USA, France and Austria.
We have already obtained some promising results, though on relatively small datasets”, says Irene Gu. “Now we are conducting further in-depth research, where more students and researchers from Chalmers participate in close collaboration with Sahlgrenska University Hospital.”
Text: Yvonne Jonsson, Communication officer at the Dept. of Electrical Engineering, E2, Chalmers