Ruipeng Li ^1* Yuehui Liao ² Yueqi Huang ³ Xiaofei Ma ² Guohua Zhao ⁴ Yanbin Wang ¹ Chen Song ¹

• ¹ Hangzhou Third People’s Hospital, Hangzhou, China
• ² Zhejiang Chinese Medical University, Hangzhou, Zhejiang Province, China
• ³ Hangzhou Seventh People’s Hospital, Hangzhou, Jiangsu Province, China
• ⁴ First Affiliated Hospital of Zhengzhou University, Zhengzhou, Henan Province, China

Glioma segmentation is crucial for diagnostic decisions, monitoring disease progression, and planning surgical interventions. However, it faces significant challenges due to the substantial heterogeneity within gliomas and imbalanced distributions of regions. This paper presents the DeepGlioSeg network as a solution to these challenges. The network architecture is designed in a U-shape and includes skip connections to ensure the continuous integration of contextual information. The DeepGlioSeg model comprises two critical components. First, it employs a CTPC (CNN-Transformer Parallel Combination) module, which consists of parallel branches of CNN and Transformer networks. This module facilitates the fusion of local and global features in glioma images by enabling interactions between these two branches. Second, DeepGlioSeg estimates the probability of each region by calculating the ratio of the pixels in the tumor region to those in the background region. It assigns greater weight to regions with lower probabilities, thereby increasing the model's focus on accurately segmenting the tumor region. Finally, DeepGlioSeg incorporates test-time augmentation (TTA) and volumeconstrained (VC) post-processing techniques to derive the final segmentation results from glioma MRI data. Comprehensive experiments were performed on three publicly available datasets and one privately owned clinical dataset. The results demonstrate that DeepGlioSeg is effective and superior to other state-of-the-art segmentation methods. The source code for this work has been made publicly available at https://github.com/smallboy-code/Brain-tumor-segmentation.