Cell nuclei segmentation is a fundamental task in computational pathology, crucial for characterizing the tumor microenvironment and supporting prognosis prediction. The recently proposed Segment Anything Model (SAM) has demonstrated strong general-purpose image segmentation capabilities. However, its direct application to histopathological images faces notable limitations, including difficulty in resolving overlapping nuclei, sensitivity to background artifacts, and inconsistent performance under diverse staining conditions. To address these challenges, we propose Nu-SAM, a novel framework built upon SAM and tailored for densely packed nuclei segmentation in pathological images. Nu-SAM integrates three key enhancements: an Adapter-based SAM2 Encoder (ABSE) for efficient fine-tuning with minimal parameter updates; a High-Low Frequency Fusion Module (HLFM) that utilizes Fast Fourier Transform to extract and fuse high-frequency boundary details and low-frequency contextual information; and a Channel and Spatial Attention Residual (CSAR) module to enhance discriminative feature learning by capturing spatial and channel-wise dependencies. Extensive experimental results demonstrate that Nu-SAM outperforms state-of-the-art methods, delivering superior segmentation accuracy and robustness in densely populated cellular environments with marked staining variability. In conclusion, Nu-SAM achieves excellent performance in cell nuclei segmentation and offers a reliable solution for DAPI-stained images analysis. This work highlights Nu-SAM’s potential as a reliable and generalizable solution for automated histopathological image analysis, thereby facilitating advances in computational pathology and precision oncology.