Using the variability of the human cortical folding pattern to benchmark unsupervised learning – FOLDDICO

The main goal of the project is the design of a deep learning approach that can generate interesting regional representations of cortical folding variability (Beta-VAE, self-supervised learning, GAN, etc...)

Reconstructions from the latent spaces inferred by beta-VAEs highlight the patterns noticed by psychiatrists in the cingulate region of the cortex.

Working on several primate species in parallel will facilitate the design of adequate deep learning architectures.

Anomaly detectors will allow to highlight developmental phenomena at the origin of weaknesses.

1. Balzeau, A., & Mangin, J. F. (2021). What Are the Synergies between Paleoanthropology and Brain Imaging?. Symmetry, 13(10), 1974.
2. Borne, L., Riviére, D., Cachia, A., Roca, P., Mellerio, C., Oppenheim, C., & Mangin, J. F. (2021). Automatic recognition of specific local cortical folding patterns. NeuroImage, 118208.

1. Guillon, L., Cagna, B., Dufumier, B., Chavas, J., Rivière, D., & Mangin, J. F. (2021, September). Detection of abnormal folding patterns with unsupervised deep generative models. In International Workshop on Machine Learning in Clinical Neuroimaging (pp. 63-72). Springer, Cham.

There are more than seven billion humans alive today, who have both unique fingerprints and unique cortical folding patterns. Nobody knows whether the variability in these cortical folding patterns has a meaning, or even where these patterns, which appear in utero, come from. In mainstream brain mapping methodologies, this variability is treated as noise, and is cancelled out as far as possible when warping each brain to a template. However, some research groups consider that cortical folding could become a useful proxy for cortical architecture, as evidenced by the distinctive patterns observed in some developmental diseases like epilepsy.

Current efforts towards a better exploitation of cortical folding variability are hindered by the mismatch between the simplistic folding template found in anatomical textbooks and the incompatible configurations frequently observed in the general population. Today, millions of brains have been digitized using magnetic resonance imaging (MRI), thus creating a valuable opportunity. The aim of FoldDico is to turn this opportunity into an attractive benchmark for unsupervised machine learning, through the inference of a dictionary of the cortical folding patterns observed in humans. This dictionary, inferred from the general population, will be used to mine large databases of psychiatric patients to enable the detection of unusual folding patterns indicating abnormal developmental events. It will also enable us to create a pioneering brain mapping methodology, imposing local compatibility of the folding pattern when comparing subjects. A folding dictionary of several great ape species will also be inferred to tackle evolutionary issues.

In addition to performing our own dictionary inference, we will design a framework leveraging public and private databases to allow the community to benchmark alternative approaches. The quantitative benchmarks will build upon the capacities of the dictionaries to detect specific pattern distributions in developmental diseases and to explain the variability of the functional architecture of the cortex in the general population (inferred from functional and diffusion MRI).