MRI segmentation of head and neck tumors using hybrid attention mechanism and dense dilated spatial pyramid pooling

Head and neck cancer (HNC) involves anatomically intricate regions where precise target delineation is essential for radiotherapy. The superior soft-tissue contrast of MRI provides clearer boundary visualization compared with computed tomography (CT), enabling tighter margins and supporting daily plan adaptation in online adaptive radiotherapy. However, despite the advances of U-Net-based deep learning models, tumor segmentation in HNC remains challenging due to ill-defined borders and heterogeneous intensity patterns, which limit feature extraction and compromise small-lesion recognition.

To overcome the limitations of traditional approaches, this study proposes an improved SCDU-Net model that integrates collaborative spatial-channel attention mechanisms with densely connected atrous spatial pyramid pooling techniques, aiming to significantly enhance the segmentation accuracy and robustness of head and neck tumor MRI images.

SCDU-Net integrates two modified modules to improve segmentation capability. The model incorporates a spatial-channel dual attention module (SC) in the decoding pathway, which strengthens critical tumor feature expression through adaptive channel weight adjustment mechanisms, while capturing long-range spatial dependencies using coordinate axis attention to improve localization accuracy of small target lesions. Additionally, the network embeds a densely connected atrous spatial pyramid pooling module dense atrous spatial pyramid pooling (DenseASPP) in the bottleneck layer, which enhances edge contour detail perception through multi-scale receptive field fusion strategies, improving the network's segmentation performance.

Our proposed model is evaluated on the publicly available HNTS-MRG2024 dataset, showing promising results compared to existing approaches.

The results indicate that by integrating two modules, our method performs spatial and channel feature recalibration and multi-scale contextual modeling within the deep neural network, yielding more accurate and promising head and neck tumor segmentation with potential to assist physicians in diagnosis.