Computer Vision News - March 2023

41 Chiara Mauri which not only yields accurate predictions, but is also inherently interpretable, and simple and fast to use.  I first demonstrated the effectiveness of the method by using age and gender prediction tasks: I showed that the method achieves prediction performances that are competitive with state-of-the-art techniques, especially whenthesamplesize isatmostof a fewthousand subjects, which is the typical scenario in many neuroimaging applications. I then employed the method to perform automatic diagnosis of multiple sclerosis patients.   A huge advantage of this method is that, in addition to being accurate, it is also inherently interpretable. In fact, it automatically provides maps displaying the morphological effect of the variable of interest on the brain, which are straightforward to interpret. Figure 1 shows an example of these maps obtained for age prediction: they mostly highlight areas around the ventricles and in the cortex, expressing therefore ventricles enlargement and cortical thinning, which are known morphological effects of brain aging.  This prediction method also allows to answer counterfactual questions, such as “ How will my brain look like when I am 80 years old? ”. Figure 2 shows the real brain image of a 47 years old person, together with the same person’s brain, artificially aged to 80 years old. We can see that the model is able to encode typical aging patterns, such as ventricles enlargement.  Therefore, this prediction method that I developed has the advantage of being both accurate and interpretable, but let’s not forget about its simplicity and speed: training the method on a few hundred of subjects for example takes only a few minutes (on a CPU)!