*4.2. Endocast Side*

Variation is an important concept in paleoanthropology. Paleoneurological approaches try to identify as precisely as possible intraspecific variations, as well as diagnostic features, between species. In turn, the initial mainstream paradigm in brain mapping involved canceling out morphological variability to allow comparative analysis of the functional maps across subjects and experiments. Neuroimaging, however, has widened its scope during the last decades to the modeling of intersubject variability, in order to tackle the discovery of biomarkers of pathology or the stratification of populations of patients. Furthermore, neuroimaging is now widely used to understand brain development and to compare primate species. It is time to consider cross-fertilization with paleoneurology, which has evolved in a niche built upon geometric morphometrics, which has prevented synergies. Broadening our knowledge of brain variability in our species by including a long time dimension will be of grea<sup>t</sup> help in defining the brain anatomy of *H. sapiens*. It also opens up perspectives for understanding how our brain works.

One original and exciting perspective will be to reconstruct a fossil hominin brain. A recent study [108] was the first to attempt the reconstruction of a *H. neanderthalensis* brain by deforming a population average brain for modern humans into the shape of the endocast of a reconstituted Neandertal. However, this approach does not consider the differences in brain structure between these species, such as those that we documented [103]. The different approaches detailed here, aiming at the collection of better information on the brain/endocast correspondence in living humans, developing new tools of automatic determination of the sulci on the endocasts, and enlarging our knowledge of fossil hominin variation thanks to a better availability of high-quality endocranial surfaces, will make it possible to obtain more satisfactory results.

Modern Artificial Intelligence could even play a role in the cross-fertilization between paleoneurology and neuroimaging. Provided that dedicated MRI sequences can deliver consistent proxies of endocasts on a large scale, deep learning could be trained to transform endocasts into standard representations of the cortical surface used in the mainstream neuroimaging field. Transfer learning could be tested on extant non-human primates and applied to extinct species in case of success.
