DAG-UNet: Dissimilarity-Aware Global Context Guided Lightweight UNet for Medical Image Segmentation

Convolutional Neural Networks (CNNs) have shown promising results in the field of medical image segmentation, but often struggle to capture global context. While transformer-based models can address this limitation, they come with high computational costs. To tackle the challenges of high computational complexity and the limitations of CNNs, we propose DAG-UNet. The DAG-UNet reduces channel-wise redundancy and enhances feature diversity by introducing a similarity-based strategy to select dissimilar feature channels while efficiently capturing global context through multilayer perceptron. The DAG-UNet is also lightweight, with only 0.72M parameters and 0.62G FLOPs. We evaluate the performance of DAG-UNet on Speech MRI and ultrasound datasets. It achieves competitive performance on Speech MRI and outperforms other models on the ultrasound dataset. The code is available at: https://github.com/Yhe9718/DAG-UNet.