Multi-resolution based dual-channel UNet with cross clique for medical image dense prediction

Multi-resolution strategy is a powerful way for the tasks of medical image dense prediction. However, previous approaches generally utilize multi-resolution input images only with the original single-resolution labels for model training, and therefore do not make full use of the multi-resolution supervision information to further improve model performance. To this end, a novel multi-resolution based Dual-Channel UNet with Cross Clique, named DCCC-UNet, is proposed for medical image dense prediction. It generates multiple low-resolution images together with their corresponding labels losslessly from the original high-resolution images for a dual-channel UNet. This specially designed dual-channel UNet consists of a standard UNet and a grouped UNet to learn feature representations from the high- and low-resolution data, respectively. The additional labels of lowresolution images thus further promote model training. Moreover, a new cross clique block is developed to be embedded between two UNet channels, which strengthens the feature interaction not only between the two UNet channels, but also among different convolution groups in the layers of the grouped UNet. The outputs from the two UNet channels are then fused by using the spatial attention to produce the dense prediction results. Extensive experiments are conducted on two public datasets for image segmentation and reconstruction tasks, respectively. The experimental results indicate that the proposed DCCC-UNet achieves the best performance on both datasets, suggesting its effectiveness for medical image dense prediction task.