LeViT-UNet: Make Faster Encoders with Transformer for Medical Image Segmentation

Medical image segmentation plays an essential role in developing computer-assisted diagnosis and treatment systems, yet it still faces numerous challenges. In the past few years, Convolutional Neural Networks (CNNs) have been successfully applied to the task of medical image segmentation. Regrettably, due to the locality of convolution operations, these CNN-based architectures have their limitations in learning global context information in images, which might be crucial to the success of medical image segmentation. Meanwhile, the vision Transformer (ViT) architectures own the remarkable ability to extract long-range semantic features with the shortcoming of their computation complexity. To make medical image segmentation more efficient and accurate, we present a novel light-weight architecture named LeViT-UNet, which integrates multi-stage Transformer blocks in the encoder via LeViT, aiming to explore the effectiveness of fusion between local and global features together. Our experiments on two challenging segmentation benchmarks indicate that the proposed LeViT-UNet achieved competitive performance compared with various state-of-the-art methods in terms of efficiency and accuracy, suggesting that LeViT can be a faster feature encoder for medical images segmentation. LeViT-UNet-384, for instance, achieves Dice similarity coefficient (DSC) of 78.53% and 90.32% with a segmentation speed of 85 frames per second (FPS) in the Synapse and ACDC datasets, respectively. Therefore, the proposed architecture could be beneficial for prospective clinic trials conducted by the radiologists. Our source codes are publicly available at https://github.com/apple1986/LeViT_UNet.