Enhancing perihematomal edema segmentation: integrating prior knowledge with deep learning for enhanced accuracy and interpretability.

Background: Perihematomal edema (PHE) in spontaneous intracranial hemorrhage can lead to intracranial hypertension, cerebral hernia, and even death; therefore, accurate segmentation of PHE has some value in determining medical management strategies. Despite the rapid advancements in deep learning (DL), the gray-scale overlap and indistinct boundaries between PHE and surrounding tissues continue to pose challenges for segmentation. The aim of this study was to enhance the efficiency and reliability of clinical diagnosis by improving the accuracy of PHE segmentation. Methods: Given that prior knowledge of PHE provides essential location information and feature variations, we integrate this knowledge with DL methods, which are expected to significantly enhance the accuracy and robustness of segmentation. In this work, we propose a network called the Perihematomal Edema Synergistic Enhancement Network (PESE-Net) for PHE segmentation. In the PESE-Net, we first screen PHE-associated slices in spontaneous intracranial hemorrhage images using the slice similarity-based PHE-associated slice generation method. Then, we employ a novel feature weighting strategy to synergistically fuse the overall change features and spatial information of PHE, thereby enhancing the overall intracranial expression features of PHE. Results: We evaluated the proposed PESE-Net and compared it with several state-of-the-art methods. The experimental results demonstrated that our method performs well across most evaluation metrics and achieves the best performance on the metric of relative volume difference. Conclusions: Our proposed PESE-Net leverages the contextual relations of PHE between consecutive slices as a priori knowledge for PHE segmentation, achieving effective segmentation and stable volume estimation of PHE regions.