APCP-NET: Aggregated Parallel Cross-Scale Pyramid Network for CMR Segmentation

Cardiac magnetic resonance (CMR) segmentation has been a glaring medical imaging issue for the past decades. The existing convolutional networks used for CMR segmentation lack adequate communication between semantic and spatial information. In this paper, we propose an aggregated parallel cross-scale pyramid network (APCP-Net) for CMR segmentation, which adopts parallel cross-scale spatial pyramid to enhance the ability of multi-scale feature fusion. To alleviate the computational complexity of the network without losing performance and increasing time consumption, we present a convolution decomposition strategy which decomposes a regular convolutional layer into several lightweight ones. Furthermore, deep aggregation is used among parallel spatial pyramids to promote the representation ability of our APCP-Net. We evaluate our APCP-Net on the segmentation dataset of ACDC 2017 Challenge. Experimental results demonstrate that our APCP-Net is comparable to the state-of-the-art methods in the case where the parameter amount is 54 times less than the existing methods.