Medical priors-guided feature learning network on multimodal imaging raw data for brain tumor segmentation

Abstract

Mainstream brain tumor segmentation methods require skull stripping, which can inadvertently remove adjacent tumor lesions and reduce accuracy. To address this, we propose MPGNet, which directly uses raw multimodal imaging data for segmentation. Guided by medical prior information, it effectively avoids skull interference and improves accuracy. Specifically, to alleviate skull interference and misidentification, we design a relevant graph aggregation (RGA) module that enhances feature representations by leveraging the structural characteristics of the brain. Then, to reduce confusion among different regions in the prediction results, we define a prior density loss (PDL) function using brain tumor density information from multimodal imaging. Finally, to evaluate our method, we collect skull-stripped brain tumor segmentation challenge (BRATS) data, their corresponding Cancer Genome Atlas (TCGA) raw data, and actual clinical raw data annotated by experienced radiologists. Our experiments demonstrate that MPGNet is effective at preserving tumor integrity compared to other state-of-the-art brain tumor segmentation methods that require skull stripping, improving the Dice similarity coefficient by 4.27%. Additionally, when all models are trained and tested with raw data, MPGNet outperforms the best existing model by 1.05% Dice, showcasing superior performance in handling skull interference.