Abstract:
Recent researches on medical image segmentation resort to the combination of natural image segmentation models and medical domain knowledge. However, prior methods only focus on single image segmentation or 3D convolutional operation based volume segmentation, and overlook the spatial correlations of inter-slice and temporal correlations of the inter-sequence in DCE-MRI images. In this paper, we propose a novel end-to-end temporal-spatial graph attention network (TSGAN), which precisely segments tumor of 4D (volume space, time) DCE-MRI images by conjointly exploiting the spatial contextual dependency of inter-slice and temporal contextual dependency of inter-sequence. Specially, we design a graph temporal attention module to integrate the temporal-spatial representations hidden in 4D data into deep segmentation. The spatial dependency is learnt by graph attention operation, which attends over its neighbourhoods' features for each vertex. Meanwhile, the spatial representations learnt by the graph attention layer are combined with the temporal representations by a temporal attention operator. Then the temporal dependency is exploited by spreading on the graph. We also design a tumour structural similarity (TSS) loss used to exploit the tumour structural dependency and enhance inter-voxel similarity within the same tissue for segmentation. We demonstrate that the proposed model outperforms recent state-of-the-art methods through comprehensive experiments.
