Accurate nuclear segmentation in histopathological images is critical for disease diagnosis and treatment planning, yet remains challenging due to nuclear crowding, overlapping, and morphological variations. To address these issues, we present SRL-UNet, an advanced segmentation framework that augments the attention-based U-Net architecture with 2D-Selective-Scan (SS2D) modules and residual connections in the encoder. This design enhances contextual representation and ensures efficient feature propagation in complex pathological contexts. Additionally, we propose the logarithmic Dice-BCE (LDB) loss function, which applies a logarithmic transformation to the standard Dice coefficient before integrating it with binary cross-entropy. The LDB loss leverages this transformation to heighten sensitivity to subtle overlap differences and dynamically adjusts class weights to mitigate class imbalance. To enhance interpretability, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) to demonstrate the SS2D module’s capability to focus on nuclear morphology despite occlusion. Experiments on the publicly available MoNuSeg benchmark dataset (H&E stained tissue samples from multiple organs) show that SRL-UNet performs superior in segmentation performance compared to current mainstream algorithms, achieving a Dice coefficient of 85.08% and an Intersection-over-Union (IoU) of 74.30%. These results demonstrate the robustness and generalizability of SRL-UNet for nuclear segmentation in diverse histopathological images.