MobileUNETR+: An Efficient Medical Image Segmentation Network Integrating Self-Attention and Dual-Pathway Strategies

Accurate medical image segmentation is essential for clinical diagnosis and treatment planning, especially in the early detection of skin cancer. Although U-Net-based architectures have demonstrated strong performance, their reliance on deep convolutional hierarchies and skip connections often result in high computational overhead and limited capacity to model global context. Transformer-based approaches address this issue through self-attention mechanisms, but their quadratic complexity restricts their applicability in real-time and resource- constrained environments. To tackle these challenges, we propose MobileUNETR+, a lightweight hybrid segmentation network that integrates an Efficient Non-Local Transformer Block (ENLTB) with linear complexity and a dual-path encoder composed of depthwise separable and dilated convolutions. This design enables efficient fusion of fine-grained local features and long-range semantic dependencies, while maintaining low model complexity. Extensive evaluations on multiple public skin lesion segmentation datasets demonstrate that MobileUNETR+ consistently outperforms state-of-the-art methods in segmentation accuracy and model efficiency. With significantly fewer parameters and faster inference speed, MobileUNETR+ is well-suited for real-time clinical applications and edge-device deployment.