Multimodal State-Space Fusion for Small-Lesion Segmentation with Missing Input Modalities

Abstract

The accurate segmentation of small lesions in medical imaging is a critical prerequisite for early diagnosis, treatment planning, and prognostic evaluation across numerous pathologies. While multimodal imaging provides complementary anatomical and physiological information that significantly enhances segmentation performance, clinical realities frequently dictate that one or more modalities may be missing due to acquisition time constraints, patient motion artifacts, or resource limitations. This paper proposes a novel framework centered on multimodal state-space fusion designed explicitly to address the dual challenges of delineating minute structures and maintaining robustness when subjected to missing input modalities. By leveraging the continuous-time dynamics and long-range dependency modeling capabilities of state-space models, the proposed architecture maps disparate imaging modalities into a unified, high-dimensional hidden state space. This mapping facilitates dynamic feature aggregation and compensates for absent data streams through a learned consensus representation. Comprehensive theoretical analysis and extensive evaluations on standardized lesion segmentation benchmarks demonstrate that the state-space fusion paradigm fundamentally outperforms traditional convolutional and transformer-based approaches in preserving high-frequency spatial details essential for small-lesion recovery. Furthermore, under various missing modality scenarios, the framework exhibits exceptional resilience, retaining high segmentation fidelity without requiring complex generative imputation techniques. The findings establish a robust foundation for developing deployable clinical segmentation systems that operate reliably under imperfect data conditions.