Classification of Brain Tissues in Hyperspectral Images Using Vision Transformers

Hyperspectral (HS) imaging is a non-invasive technique able to identify the components of a sample by its reflectance properties. Nonetheless, the classification task of HS images is a complex problem, so advanced techniques as deep learning (DL) have been proposed in the literature to perform this task. Although these methods have generated competitive results, there are still open problems, such as the limitation of using global information and the high computational cost. Recently, vision transformers (ViT) have emerged as alternatives in some DL problems with excellent results on synthetic data. In this work, we propose to use ViTs to classify HS images of in-vivo real brain tissue to detect Glioblastoma-type tumor tissue. The classification results demonstrate that ViTs generate high performance predictions, with an overall average accuracy of over 99% and 86% for intra-patient and inter-patient training approaches, respectively.