Mamba-Based Multimodal Fusion for Robust Medical Image Segmentation and Lesion Delineation

Abstract

Medical image segmentation and lesion delineation are essential tasks in diagnostic imaging, oncology, and image-guided intervention, but their clinical utility depends on robustness across imaging modalities, acquisition sites, patient populations, and lesion types. Recent advances in state space modeling, particularly the Mamba selective scan mechanism, offer a promising alternative to convolutional and transformer-based segmentation pipelines by combining long-range dependency modeling with improved computational scaling. When embedded within multimodal fusion architectures, Mamba-based models can integrate complementary imaging sequences while preserving boundary detail and reducing memory demands. However, the design of such systems extends well beyond model architecture. This paper presents a system-level examination of Mamba-based multimodal fusion for segmentation and lesion delineation, focusing on structural trade-offs, robustness, uncertainty, data governance, fairness, infrastructure, deployment, sustainability, and policy implications. We argue that clinical value emerges not from segmentation accuracy alone but from the alignment of model design with institutional workflows, regulatory requirements, and sociotechnical safeguards. The discussion integrates architectural analysis with cross-domain comparisons and forward-looking perspectives on federated evaluation, resource-aware deployment, and accountable clinical integration. The result is a comprehensive framework for developing and deploying multimodal segmentation systems that are technically capable, operationally sustainable, and ethically defensible.