Multi-domain cotraining for tissue segmentation in fixed and fresh brain tissue using Mueller polarimetry (Conference Presentation)

Abstract

Delineating the boundary of a tumors from healthy brain tissue is a challenging task in neurosurgery. Mueller polarimetry imaging promises to visualise and segment these borders in real-time, based on optical properties correlated with the directionality of densely packed white-matter fiber-bundles. In prior work, we demonstrated deep-learning methods leveraging Mueller polarimetry outperformed traditional approaches with similar segmentation tasks. However, formalin-fixation vs. fresh sample tissue and differences of human vs. animal brain tissue properties may hinder the direct applicability to neurosurgical scenarios. To overcome this potential limitation, we propose a learning-based strategy by jointly training on augmented multi-domain data together with model fine-tuning to improve tissue segmentation.