Cross Attention Densely Connected Networks for Multiple Sclerosis Lesion Segmentation

Automatic lesion segmentation on conventional magnetic resonance imaging is an essential component in disease diagnosis, assessment, and follow-up. Recently, extensive deep neural networks have been designed for automatic lesion segmentation. However, these approaches are not easy to further optimize owing to the poor interpretability. In this paper, we present a novel cross attention densely-connected network(CA-DCN) for multiple sclerosis lesion segmentation, which integrates attention mechanism into the encoder-decoder architecture. Aiming for further improving the performance of the model, we propose a comprehensive cross attention mechanism module by combining the characteristics of spatial and channel domains. Our method is evaluated on the public International Symposium on Biomedical Imaging (ISBI) 2015 multiple sclerosis segmentation challenge. At the time of submission, our method was amongst the top performing solution.