Abstract:

When the feature of stroke lesions is non-distinct,and the boundary between the lesions and the healthy brain tissue is difficult to distinguish,the segmentation model based on the self-attention mechanism is prone to generate a wrong attention coefficient map of the focus area,which affects the segmentation performance.To solve this problem,based on the global-attention-upsample attention U-Net (GAU-A-UNet),we propose a primary-auxiliary path attention compensation network (PAPAC-Net).The primary path network is responsible for accurate lesion segmentation and outputting the segmentation results while the auxiliary path network generates a tolerant auxiliary attention compensation coefficient to compensate for the primary path network’s potential attention coefficient map errors.Two compound loss functions are also proposed to realize the different functions for the primary and auxiliary path networks.Experimental results show that our GAU-A-UNet and PAPAC-Net both have a significant improvement in segmentation performance,which proves the effectiveness of our method.

Key words: stroke, segmentation, deep learning, convolutional neural networks, attention