​The surface of the brain is divided into numerous convolutions, called gyri, bounded by folds, called sulci. These folds in the surface of the cortex are one of the most striking features of brain anatomy. Their variability is such that our "brain print" is unique, like our fingerprint. Because of its great variability, this feature is still poorly understood today. Several studies have shown that unusual folding patterns are often linked to abnormal developments that can lead to syndromes such as epilepsy or schizophrenia. Moreover, a close link has been demonstrated between the shape of certain folds and cognitive specificities such as manual laterality or a reading score. Fold patterns thus seem to be signatures of the functional organization of the brain of each individual.
However, the study of cortical folds requires a high level of expertise that few neuroanatomists currently possess, and manual classification of local sulcal patterns is a time-consuming task. Indeed, 3D visualization of the folds helps experts to identify the different patterns but the variability of the folds is so high that distinguishing between these patterns sometimes requires the definition of complex criteria, making manual classification difficult and unreliable. Automation of fold pattern recognition is essential to allow for expansion and confirmation of these studies, while using larger databases that would lead to a better understanding of the subtle links between sulcal shapes and functional architecture, especially when analyzing rare profiles.

In this paper, three automatic fold pattern classification algorithms were used to enable the extension and confirmation of morphological studies on such large databases: a Support Vector Machine (SVM) classifier, a Scoring by Non-local Image Patch Estimator (SNIPE) metric, and a 3D Convolutional Neural Network, CNN. These methods were tested on two types of patterns for which there is currently no approach to automatically identify them: two patterns in the Anterior Cingulate Cortex (ACC), both of which are equally represented in the general population, and the Power Button Sign (PBS) (Figure), a particularly rare pattern related to epilepsy arising from motor regions. The three proposed models achieve balanced accuracies of about 80% for ACC pattern classification and 60% for PBS classification. The CNN-based model is more attractive for ACC pattern classification due to its fast execution time. However, the SVM and SNIPE based models are more efficient in handling unbalanced problems such as PBS recognition.

In conclusion, none of the proposed models is better than the other two because their performance in identifying individual patterns is equivalent, and although the fastest model is the CNN-based model, it is also the least efficient in finding rare patterns. Artificial intelligence has made significant progress, especially in the field of computer vision, revolutionizing our daily lives. It will certainly allow us one day to decipher the lines of the brain.

Contact : Jean-François Mangin jean-francois.mangin@cea.fr