R-DWSAxNet: Residual Depthwise Separable Convolution with Axial Attention for efficient multi-modal medical image segmentation

Accurate and efficient medical image segmentation remains a central challenge in computer-assisted diagnosis, particularly when dealing with heterogeneous imaging modalities and limited computational resources. Traditional convolutional models, while effective in capturing local spatial patterns, often struggle to model long-range dependencies and require substantial processing power. To address these limitations, we introduce R-DWSAxNet, a lightweight hybrid architecture that integrates Residual Depthwise Separable Convolutions, Axial Attention, Squeeze-and-Excitation mechanisms, and an efficient ASPP-Lite module. This design significantly reduces computational complexity while enhancing contextual feature extraction and boundary precision. The proposed model is evaluated on four diverse medical imaging benchmarks CVC-ClinicDB, PH2, BraTS2018, and Montgomery. Experimental results show that R-DWSAxNet consistently outperforms classical and contemporary U-Net variants, achieving superior Dice coefficient and IoU scores across all datasets. Despite its strong performance, the architecture remains highly efficient, requiring only 8.15M parameters and reduced GFLOPs, making it suitable for real-time or resource-constrained clinical environments. These findings highlight the potential of R-DWSAxNet as a robust, scalable, and computationally efficient solution for multi-modal medical image segmentation.