A Lightweight Hybrid CNN-Mamba Model with Dual Attention for Efficient Medical Image Segmentation

Abstract

In the medical field, early diagnosis plays a crucial role and directly influences treatment methods. Particularly in image processing, identifying harmful cells is vital for health. With advancing technology, deep learning models have been widely applied in biomedical image analysis and segmentation to enhance efficiency and accuracy. CNN-based and Transformer-based models have been extensively researched and utilized. However, CNN models often struggle to capture long-range dependencies, while Transformers face challenges with quadratic computational complexity. To address these issues, a recent method called Mamba has emerged, by the building upon Mamba-based models, we propose a novel architecture, DA-Lite Mamba, incorporating Double Attention CNN and Mamba, ConvBlock, Bottleneck Block, Decoder Block and Map Reduce and Interpolation Block. Our work introduces two new feature extraction architectures: Double Attention CNN and Mamba and Bottleneck Block, which effectively leverage Depthwise Convolution and VSS Block to balance local and global information extraction, enhancing comprehensive feature representation while minimizing computational parameters. Additionally, we propose a novel loss calculation method based on the Matthews Correlation Coefficient (MCC) Loss to improve the convergence ability of deep learning networks. The proposed model is evaluated on two benchmark datasets: Sunnybrook Cardiac Dataset (SCD) and Data Science Bowl 2018 (DSB 2018). Experimental results demonstrate superior performance compared to previous CNN- and Transformer-based models across various evaluation metrics.