Adaptive Lightweight Framework for the Brain MRI Segmentation Task

Medical image segmentation has always been one of the most challenging tasks in the field of computer vision. The requirement for segmentation accuracy makes it difficult to train models with an increase in parameters. In this paper, we propose a novel lightweight network. This network is based on the U-Net network structure, but discards the pooling layer, which may cause a serious decrease in accuracy during downsampling, and instead uses a convolutional layer. We also use lightweight deep convolution as part of the feature extraction. At the same time, in order to maintain representational power, we remove the nonlinearity in the narrow layer. In order to study the generalization of this method, we also experimented with different attention mechanisms in the Brain MRI segmentation task. The experimental results show that this model can not only greatly reduce the number of parameters of the original Attention U-net, but also perform better in the Brain MRI segmentation task.