Feature Alignment in Vision Mamba to Resolve Domain Shift of Mobile Medical Devices

Integrating deep learning models into mobile medical devices has led to remarkable growth in medical diagnosis and healthcare applications. In particular, the recently emerged vision mamba (ViM) has been attracting attention because it shows better accuracy than mobile-specific models and is very suitable for mobile environments where computational efficiency and memory management are important. However, there is still a lack of research on domain shift, which hinders its application in the field. To address the domain shift issue in this study, we design and integrate a squeeze and excitation layer-based auxiliary network (SEAN) into the ViM block to perform a feature alignment strategy. This enables robust performance under different domains, and in experiments on Camelyon17, a medical dataset, the proposed method achieves 0.6% and 3.2% performance improvement on ViM-Tiny and ViM-Small models, respectively. Furthermore, we show that unreliable data can be extracted using the designed SEAN, which is expected to enable more sophisticated medical diagnosis in fields where resource constraints are evident.