A U-Shaped Architecture Based on Attention Mechanism
Chang Ming Lu, Lan Lin, Minyang Xu · 2023
With the rapid development of computer and artificial intelligence, image segmentation technology based on deep learning plays an increasingly important role in the medical field. In this paper, we propose the attention mechanism-based U-Net++, a new and more powerful medical image segmentation architecture. Our architecture is essentially an attention-based encoder-decoder network that uses dense skip connection to capture features at different levels, by introducing channel attention and spatial attention, the neural network pays more attention to the pixel regions that are decisive for classification and ignores the insignificant regions, so as to handle the distribution relationship of the feature map channels. In addition, we design a hybrid loss function that integrates the cross-entropy loss, Dice loss, and Focal loss to focus more on the part which is difficult to divide of the medical image dataset. We evaluate the attention-based U-Net++ compared to U-Net and UNet++ architectures in a medical image segmentation task for skin lesion datasets. Our experiments show that the average IoU gain of the U-Net++ networks based on the attention mechanism exceeds that of general U-Net and UNet++.