<![CDATA[Diabetic retinopathy is a major microvascular complication of diabetes and a leading cause of vision loss among the working-age population. Microaneurysms (MAs), as the earliest clinical indicators of disease progression, remain challenging to segment due to their small size, low contrast, and extreme class imbalance. This study proposes a lightweight hybrid U-Net architecture for microaneurysm segmentation in retinal fundus images, designed to balance detection sensitivity and computational efficiency for deployment in resource-constrained environments. The proposed architecture integrates depthwise separable convolutions for efficient local feature extraction, a Transformer-Lite bottleneck based on Linformer self-attention for global contextual modeling, and a Mamba State Space Model (SSM)–based decoder to enhance feature propagation and spatial continuity. The research contribution of this work is threefold: the introduction of an efficient hybrid U-Net combining Linformer and Mamba SSM for microaneurysm segmentation; a deployment-oriented evaluation protocol that explicitly distinguishes patch-level learning behavior from full-image reconstruction performance; and a transparent analysis of false positive behavior under extreme background dominance. Experiments were conducted on the IDRiD dataset, consisting of 81 retinal images, using patient-level data splitting prior to patch extraction to prevent data leakage. The results indicate that while patch-level evaluation demonstrates effective lesion-centric learning, deployment-realistic full-image evaluation reveals a notable performance degradation caused by false positive accumulation in extensive background regions. Nevertheless, the model maintains high recall, indicating preserved lesion sensitivity. These findings suggest that lightweight architectural design can deliver meaningful performance and is well suited for screening-oriented decision-support systems that prioritize efficiency and sensitivity.]]>
