Focus on… the diagnostic by computer of brain lesion

on the May 24, 2018

From a set of quantitative MRI images, Emmanuel Barbier’s (Grenoble Institute of Neurosciences) and Florence Forbes’ (Inria / Jean Kutzmann Laboratory) teams have succeeded in locating and diagnosing several types of brain tumors completely automatically. This new "machine learning" approach is a contribution to the development of radiology of the future.
MRI, or magnetic resonance imaging, is the reference technique to obtain very detailed brain images. What is less known is that MRI is sensitive to many brain tissues’ features. To make a comparison, we can take the example of the images that satellites take from our planet: the satellites can map the temperature, the wind speed, the height of the clouds or the size of the raindrops. All these quantitative images (the temperature in ° C, the speed in km / h, the height in meter ...) describe the same territory. MRI can do the same for the brain: to obtain quantitative images that each maps a different characteristic of the brain tissues. Benjamin Lemasson, a young Inserm researcher recently returned from the United States, is working on the exploitation of these imaging protocols combining many images.

 figure diagnostic

To analyze these images, Alexis Arnaud, a PhD student, has cleverly combined various recently developed mathematical tools. At first, the computer learns the characteristics of healthy brains. Then, in diseased brain images, the computer automatically locates regions that are different from healthy tissue and then extracts their characteristics. The researcher tells the computer the diagnosis associated with each diseased brain. After learning, the computer is ready for a test! The test consists of providing the computer with unknown images from healthy brains or diseased brains. Of course, the computer does not know these test images. The computer must indicate in its unknown images if there is a lesion (in our example, a tumor) and, if so, what type of lesion is present. And the computer was a very good student: he was able to locate (100%) and diagnose (> 90%) the lesions. This new method and these results are the subject of the study published in IEEE-TMI1 .


Today, this type of quantitative images does not correspond to what is done in clinical routine in MRI services. But the study published by the Grenoble researchers shows the interest of acquiring quantitative images and raises the veil on the analytical tools that radiologists will soon have available to help them in their interpretations. In the meantime, researchers will look for the most relevant images to acquire to diagnose brain tumors as accurately as possible and will continue to develop mathematical tools.

1 Alexis Arnaud, Florence Forbes, Nicolas Coquery, Nora Collomb, Benjamin Lemasson, Emmanuel L. Barbier. Fully Automatic Lesion Localization and Characterization: Application to Brain Tumors using Multiparametric Quantitative MRI Data. IEEE Transaction on Medical Imaging. Early view.
Published on May 23, 2018