To address the poor target detection, incomplete boundary recognition, and limited robustness of existing fracture segmentation methods for slender fractures in deep metamorphic rock cores under complex interference backgrounds, an improved ADM-UNet fracture segmentation model is proposed. This method makes three improvements based on the U-Net architecture. First, the Visual State Space Block (VSS Block) is introduced to improve the continuity modeling ability for slender fractures. A multi-scale Attention Detail Enhancement Module (ADEM) is embedded in the skip connections to refine skip features, and a Deformable Residual Block (DRB) is used in the resolution reconstruction stage to reduce the loss of fracture details during feature recovery. In addition, a fracture parameter quantification method is developed based on the segmentation results. Experimental results show that ADM-UNet achieves a mean Intersection over Union, precision, and recall of 83.87%, 81.29%, and 87.24%, respectively, representing improvements of 7.63, 6.64, and 3.42 percentage points over the baseline U-Net. The proposed method enables fracture recognition and quantitative parameter characterization of deep metamorphic rock cores and provides data support for fracture development analysis and the evaluation of deep fractured metamorphic geothermal reservoirs.