Automatic segmentation of medical images is a crucial step for lesion measurement in computer-aided diagnosis. Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) are widely adopted but have limitations. To address these challenges, we propose a Frequency-enhanced Lightweight Vision Mamba Network (FMamba) for automatic medical image segmentation. Specifically, we introduce the Vision State Space (VSS) and Frequency Feature Enhancement (FFE) modules for efficient parallel feature extraction. The VSS module employs 2D-Selective-Scan (SS2D) to scan feature maps in multiple directions, effectively building long-range dependencies. At the same time, the FFE module refines the frequency domain of the feature maps, yielding enhanced global feature representations, thereby enhancing the global context awareness. Compared to UNet, our method reduces GFLOPs and Parameters by 25.99 times and 5.84 times, respectively. On the BUSI dataset, Dice and IoU scores improved by 3.25% and 3.35%, respectively. On Dataset B, improvements were 2.69% and 2.21%, respectively. Our method can effectively integrate state space model and frequency domain features, surpassing existing methods in medical image segmentation tasks.
1 September 2022 in Computers in Biology and Medicine
Accuracy segmentation of COVID-19 lesions in lung CT images can aid patient screening and diagnosis. However, the blurred, inconsistent shape and location of the lesion area poses a great challenge to this vision task. To tackle this issue, we propose a multi-scale representation learning network (MRL-Net) that integrates CNN with Transformer via two bridge unit: Dual Multi-interaction Attention (DMA) and Dual Boundary Attention (DBA). First, to obtain multi-scale local detailed feature and global contextual information, we combine low-level geometric information and high-level semantic features extracted by CNN and Transformer, respectively. Secondly, for enhanced feature representation, DMA is proposed to fuse the local detailed feature of CNN and the global context information of Transformer. Finally, DBA makes our network focus on the boundary features of the lesion, further enhancing the representational learning. Amounts of experimental results show that MRL-Net is superior to current state-of-the-art methods and achieves better COVID-19 image segmentation performance. Moreover, our network has strong robustness and generalization ability in such vision tasks as colonoscopic polyp and skin cancer segmentation.