AUTHOR=Wang Kang , Li Zeyang , Wang Haoran , Liu Siyu , Pan Mingyuan , Wang Manning , Wang Shuo , Song Zhijian 

TITLE=Improving brain tumor segmentation with anatomical prior-informed pre-training

JOURNAL=Frontiers in Medicine

VOLUME=Volume 10 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2023.1211800

DOI=10.3389/fmed.2023.1211800

ISSN=2296-858X

ABSTRACT=Accurate delineation of glioblastoma in multi-parameter magnetic resonance images is crucial for neurosurgery and follow-up treatments. Learning-based methods, especially Transformer models, have demonstrated significant potential for brain tumor segmentation. However, their effectiveness heavily relies on the availability of substantial annotated data. To mitigate the scarcity of annotated data and enhance model robustness, self-supervised learning paradigms utilizing masked autoencoders have been developed to pretrain the segmentation network. However, the anatomical knowledge of brain structures has not been integrated into the pre-training stage. In this paper, we propose an anatomical prior-informed masking strategy to optimize the pre-training of masked autoencoders, which combines data-driven reconstruction with anatomical knowledge. We investigate the probability of tumor occurrence on different brain structures which is further utilized to inform the masking process. Compared with random masking, our method enables the pre-training to concentrate on regions more pertinent to downstream segmentation. Experimental results on the BraTS21 dataset show that our proposed method outperforms the state-of-the-art self-supervised learning methods and improves brain tumor segmentation in terms of both accuracy and data-efficiency.