Brain tumors are highly dangerous diseases that can have severe implications on cognitive and physical abilities. The accurate and automated segmentation of tumor regions is of immense significance to facilitate treatment planning and pathological research. Previous segmentation methods have utilized a deep supervision strategy to optimize parameter learning. However, this strategy has mostly focused on parameter optimization during the training phase, while ignoring that knowledge from this strategy could still be advantageous during inference. Hence, our proposed method incorporates such knowledge into the forward inference of the model for feature enhancement and noisy removal. This enables our model to continuously enhance its feature representation, resulting in less erroneous segmentation. This advancement is beneficial for early and robust diagnosis of brain tumors in clinical applications.

Accurate segmentation of brain tumors is crucial for diagnostic evaluation and clinical planning. Convolutional-based and Transformer-based models have shown promising results in automatic brain tumor segmentation. In these models, a deep supervision strategy has been widely adopted for parameter optimization. As a key part of this strategy, the segmentation head is responsible for generating early segmentation in the training phase. However, although containing informative cues valuable for decoder refinement, the segmentation head is usually discarded during inference. In this work, we propose a novel approach called deep supervision guided transformer (DSGT) for brain tumor segmentation. DSGT leverages informative cues within the segmentation head to guide decoding by developing guided heads upon a Transformer-based decoder. Specifically, we first extract semantic features from the segmentation head and then design two guided modules for feature refinement and noisy removal to gener...