Abstract

In recent years, it was demonstrated that discrimination between white matter and tumor-infiltrated white matter based on optical coherence tomography (OCT) data is possible with high accuracy. However, gray matter is also present during the tumor resection and shows similar optical properties to tumor infiltration, which aggravates the tumor classification using optical coherence tomography. A semantic segmentation approach based on a convolutional neural network was applied to the problem in order to classify healthy brain tissue from tumor infiltrated brain tissue. A dataset was created, which consisted of ex vivo OCT B-scans, which were acquired by a swept-source OCT system with a central wavelength of 1300 nm. Each OCT B-scan was indirectly annotated by transforming histological labels from a corresponding H&E section onto it. The labels differentiate between white matter, gray matter and tumor infiltration. The output of the network was modeled to a Dirichlet prior distribution, which enabled the capturing of a prediction uncertainty. This approach achieved an intersection over union score of 0.72 for healthy brain tissue and 0.69 for highly tumor infiltrated brain tissue, when only confident predictions were considered.

© 2023 SPIE