Scale-Sensitive Mamba Decoding for Tiny Lesion Segmentation in Medical Imagery

Abstract

The accurate segmentation of tiny lesions in medical imagery remains a persistent challenge in computational pathology and radiology. Traditional convolutional neural networks and vision transformers often struggle with highly localized, small-scale pathological manifestations due to the inherent loss of fine-grained spatial resolution during hierarchical downsampling and the quadratic computational complexity associated with global attention mechanisms. Recently, State Space Models, particularly the Mamba architecture, have emerged as a promising alternative for modeling long-range dependencies with linear computational complexity. However, standard Mamba formulations are fundamentally sequence-oriented and tend to smooth over the high-frequency spatial details required to isolate tiny lesions. To address this limitation, this paper introduces a scale-sensitive Mamba decoding framework designed specifically for the nuanced segmentation of micro-lesions. By integrating a novel multi-scale state space projection mechanism within the decoding trajectory, our approach dynamically modulates the state transitions based on local feature variations across different resolution hierarchies. This ensures that the high-frequency structural anomalies characteristic of tiny lesions are preserved and amplified during the reconstruction phase. Comprehensive empirical evaluations across multiple imaging modalities demonstrate that the proposed architecture not only achieves superior segmentation accuracy and boundary delineation but also maintains the linear scaling efficiency characteristic of State Space Models.