Abstract:
Semantic segmentation of brain tumors is a fundamental task in medical image analysis, involving multiple MRI modalities for diagnosing patients and studying malignant tu...Show MoreMetadata
Abstract:
Semantic segmentation of brain tumors is a fundamental task in medical image analysis, involving multiple MRI modalities for diagnosing patients and studying malignant tumor progression. With the advancements in CNN and Transformer algorithms, recent 3D medical segmentation models, like VT-UNET, have achieved state-of-the-art performance in the BraTS brain tumor dataset. The majority of networks used for brain tumor semantic segmentation currently employ the U-Net architecture. While U-Net-like architectures offer rich contextual information, the process of recovering high-resolution representations from low-resolution ones may reduce the spatial accuracy of the resulting segmentation masks. In previous works, the design style of High-Resolution Network (HRNet) has been successfully applied to various 2D image segmentation tasks, such as human pose estimation, scene segmentation, and object segmentation. HRNet focuses on feature fusion between low-resolution and high-resolution representations, rather than aiming to recover high-resolution information from low-resolution ones. This approach is highly relevant for complex brain tumor segmentation tasks. Thus, in this paper, we introduce, for the first time, the utilization of the HRNet design style in the context of brain tumor segmentation. We employ the 3D UX-Net block to exchange information between feature maps of different resolutions. Extensive experiments demonstrate that on the BraTS 2021 dataset, the performance of HR-UXNet is comparable to advanced U-Net based models in brain tumor segmentation.
Date of Conference: 15-17 December 2023
Date Added to IEEE Xplore: 15 May 2024
ISBN Information: