关键词: 网络架构, 神经网络搜索, 脑肿瘤分割, 卷积神经网络, Transformer

Abstract: The significant variations in shape and size, along with blurred boundaries and complex tissue structures in glioma Magnetic Resonance Images (MRI), presented a challenging task for brain tumor segmentation. Typically, this task required researchers with specialized expertise to design complex custom network models, which was not only time-consuming but also necessitated considerable human resources. To streamline the network design process and automatically obtain the optimal network architecture, a Neural Architecture Search for multi-tissue segmentation using Convolutional and Transformer-based Networks in glioma segmentation (NASCT-Net) was proposed. This approach aimed to improve segmentation precision during the construction of a network architecture for multi-modal MRI brain tumor segmentation. Firstly, Neural Architecture Search (NAS) technology was applied to the construction of an encoder in NASCT-Net, forming stackable NAS-CNN encoding modules to automatically optimize the network structure for precise glioma segmentation. Secondly, a feature encoding module based on the Transformer was integrated into the lower layers of the encoder to improve the representation of relative positions and global information among tumor components. Lastly, a Volume-Weighted Dice Loss function (VWDiceLoss) was constructed to address the imbalance between foreground and background. NASCT-Net was compared with advanced methods such as Swin-Unet on the Brats2019 brain tumor datasets. The results show an average Dice score improvement of 0.009 and a significant increase in the average Hausdorff distance by 1.831mm. These findings validate the effectiveness of NASCT-Net in enhancing the precision of multi-tissue brain tumor segmentation.

Key words: network architecture, Network Architecture Search（NAS）, brain tumor segmentation, Convolutional Neural Network（CNN）, Transformer