Creating a depth-resolved OCT-dataset for supervised classification based on ex vivo human brain samples

Optical coherence tomography (OCT) has the potential to become an additional imaging modality for surgical guidance in the field of neurosurgery, especially when it comes to the detection of different infiltration grades of glioblastoma multiforme at the tumor border. Interpretation of the images, however, is still a big challenge. A method to create a labeled OCT dataset based on ex vivo brain samples is introduced. The tissue samples were embedded in an agarose mold giving them a distinctive shape before images were acquired with two OCT systems (spectral domain (SD) and swept source (SS) OCT) and histological sections were created and segmented by a neuropathologist. Based on the given shape, the corresponding OCT images for each histological image can be determined. The transfer of the labels from the histological images onto the OCT images was done with a non-affine image registration approach based on the tissue shape. It was demonstrated that finding OCT images of a tissue sample corresponding to segmented histological images without any color or laser marking is possible. It was also shown that the set labels can be transferred onto OCT images. The accuracy of method is 26 ± 11 pixel, which translates to 192 ± 75 μm for the SS-OCT and 94 ± 43 μm for the SD-OCT. The dataset consists of several hundred labeled OCT images, which can be used to train a classification algorithm.