Patch-wise brain age longitudinal reliability

Great news from our lab: a new article was published last month in the journal Human Brain Mapping. Congrats to Iman Beheshti as well as co-authors Olivier Potvin and Simon Duchesne for their work!

The idea behind the article is twofold. First, is that an MRI (i.e., magnetic resonance imaging) machine is – in some regards - like a high-end camera. Different brands will have different optics, lenses and so on, and therefore pictures of the same object taken with two different cameras will never end up being exactly the same; there will be variations in the clarity and contrast, for example. The same can be said for MRI machines. The differences between images from each vendor’s machines (e.g., Siemens Healthcare, Philips Medical Systems, GE Healthcare to name but a few) can range from the trivial to the consequential. Clinically, it does not matter much as radiologists are able to gloss over these differences and still arrive at a correct interpretation. For algorithms however that try to calculate measurements using the information contained in the images, this is quite a reliability puzzle. Thus, we must make sure that the tools we develop and use to analyze MRI images are robust.

The second idea in the article is that of “brain age”, an advanced machine learning and brain imaging technique that has become a hot topic in neuroimaging circles. The concept is to use machine learning to calculate the “age” of a participant’s brain by comparing his or her image to a large database of other images. The “brain age” delta (i.e., the difference between an estimated brain age and the individual's real, or chronological,m age) has been shown as a heritable metric for monitoring cognitively healthy aging and predicting risk of age-related disease. In the last decade, substantial efforts have been devoted towards the development of highly accurate brain age estimation frameworks through different brain imaging data.

We recently introduced a patch-wise technique to estimate brain age from anatomical T1-weighted MRI data which exhibited an outstanding performance, even with a limited training size. In this paper now published in Human Brain Mapping, we aimed to demonstrate the accuracty and reliability of this patch-wise technique when measuring brain age using different scanner manufacturers.

To this end, we trained our proposed patch-wise brain age estimation framework on a dataset of 100 cognitively healthy individuals from MindBoggle, and then tested on our unique dataset of MRI scans from a single, cognitively healthy volunteer that had been scanned 99 times over a period of 17 years (aged 29-46 years) at multiple sites, and therefore with a whole array of different MRI machines. Some of you will have recognized that this dataset is the SIMON, named after our Esteemed Leader ©, and the topic of another article.

Our results (Fig. 1) proved that the patch-based brain age estimation technique was extremely accurate and reliable. This opens the way for a metric like brain age to be used as a marker of decline for individuals with a neurodegenerative disorder, or to monitor the effect of interventions.

Fig.1. Influence of MRI manufacturer on brain age-delta for the single individual volunteer.

Delta: chronological age subtracted from the brain estimated age.