Training a Vision Transformer Model on Shape Signatures for Quality Control of MRI Corpus Callosum Segmentations

The corpus callosum (CC) is a vital brain structure for inter-hemispheric communication. Precise segmentation of the CC from magnetic resonance imaging is used to evaluate neurological conditions and is also associated with various individual characteristics. However, studies on quality control (QC) for segmentations are still scarce, even though they are especially necessary when automatic techniques are employed. This work explores the self-attention component of Vision Transformer (ViT) model for performing QC of CC MRI-based segmentations. Instead of feeding the ViT model with the MR image itself, the model is fed with a set of shape descriptors in several scales. This set of shape signatures is able of capturing at the same time a general shape of the CC, but also some details of the structure. Results highlight the potential of ViT models and curvature profile descriptors for the QC of CC segmentation, achieving 99% area under the curve (AUC), ensuring precision and efficiency in the evaluation of medical images through a methodology agnostic to the type of image acquisition and segmentation method.