Brain MRI, Deep Learning Aid in Gender Differentiation
Tuesday, Nov. 28, 2017
A deep learning (DL) algorithm applied to brain MRI findings helps accurately distinguish between genders, according to research presented Monday.
Hormonal abnormalities and certain diseases can make gender identification challenging. An individual may have the genotype, or genetic makeup, of one gender and the phenotype, or observable traits, of the other. For instance, a person with androgen insensitivity syndrome appears genetically male but has female genitalia.
In ambiguous cases, gender differentiation is typically accomplished through clinical examination and blood tests. An accurate imaging tool to differentiate gender would be a useful adjunct for clinicians, patients and their families, but such a test currently doesn't exist.
"Men and women have differences in their brains, but to date, no specific anatomic landmark seen by the human eye has been identified to properly distinguish between genders in sectional images," said study co-author Felipe C. Kitamura, MD, from the Federal University of São Paulo.
Dr. Kitamura and colleagues recently tested an automated method to distinguish genders using head MRI and a DL algorithm.
"The major application of this approach is to see if we can develop a different kind of biomarker using an algorithm that is capable of seeing something we are not able to see with our eyes," he said.
The researchers reviewed a total of 7,120 images from 356 patients using a convolutional neural network, a type of neural network that produces a hypothetical mathematical representation in a computer of the way the brain works. They used some of the subjects to train the algorithm, others to validate it and the rest to test it out.
When the researchers correlated the neural network findings with the subjects' genders, they found that the algorithm was 95 percent accurate at gender differentiation. Dr. Kitamura said the results show the power of DL, a subset of machine learning that provides a more in-depth analysis of data.
"With deep learning, you don't have to tell the algorithm what to look for," he said. "You give examples and it does it itself."
Despite the 95 percent accuracy of the algorithm, Dr. Kitamura said that more research is needed before it is ready for clinical use. If the results hold up over larger study groups, it could eventually have applications in the clinic.
"This approach can give us one more biomarker that will aid in the diagnosis of all those diseases that may affect gender differentiation," he said. "This could be particularly important in those cases where diseases lead to genotype/phenotype gender mismatches."
The researchers intend to look at images acquired from different MRI scanners to see if the algorithm is able to maintain its predictive power across different equipment and operators.