DCSM-Unet: Dual-path heterogeneous encoder with content-aware guidance and differential feature learning.

High-precision medical image segmentation is a prerequisite for accurate computer-aided diagnosis, yet it faces a persistent challenge in balancing fine-grained local detail extraction with global context modeling. Conventional CNN-based methods excel at analyzing local textures-critical for delineating boundaries in pathology-but suffer from limited receptive fields. Conversely, pure Transformer or state-space models (SSMs) capture the long-range dependencies needed for anatomical understanding at the cost of computational efficiency or high-frequency detail fidelity. Existing hybrid approaches often neglect the semantic heterogeneity between these architectures, leading to suboptimal feature integration. To resolve these limitations, this study introduces DCSM-Unet, a dual-path segmentation network that synergizes a ResNet-34 CNN encoder for local feature fidelity with a Mamba-based state-space encoder for global semantic consistency. Technically, we propose two core innovations: First, a content-adaptive guided fusion mechanism employs progressive optimization to generate channel-specific Spatial Weight Maps (SWM), achieving dynamic, pixel-wise calibration of heterogeneous features. Second, to mitigate detail loss during downsampling, we introduce a multi-scale differential Detail Enhancement Module (DeeM) that sharpens fine boundary representations. Extensive experiments on three public medical datasets (GlaS, MoNuSeg, and Chest X-ray) demonstrate that DCSM-Unet outperforms existing methods, achieving higher Dice, IoU, Precision, Recall and lower Hausdorff distance. The improved boundary precision 0.067-0.189 μm enhancement at 40 × magnification directly enables more accurate pathological grading in colorectal cancer diagnosis, precise nuclear morphometry measurements for tumor subtyping, and reliable lung field delineation for tuberculosis screening. These results confirm that our architecture not only effectively integrates complementary visual patterns but also provides a robust, generalized solution for diverse medical recognition tasks, ranging from precision cellular morphometry to macroscopic radiological screening.