Ensuring structural consistency during electrode winding is critical to the safety and performance of lithium-ion batteries, where the anode–cathode overhang serves as a key critical-to-safety indicator. However, CT and X-ray images collected from industrial production lines typically exhibit low contrast, strong noise, and overlapping fine structures, making accurate electrode and separator segmentation particularly challenging. Existing Segment Anything Model (SAM)-based or UNet-like methods suffer from domain shift, weak boundary fidelity, and heavy annotation requirements. This study proposes a refined SAM adaptation framework tailored for segmentation-based battery electrode analysis. The framework introduces three synergistic components: (1) encoder-level adapters and an alignment bridge for efficient domain adaptation under scarce labels; (2) a dual-branch output-enhanced decoder with multiscale fusion that preserves global context while enhancing fine-grained boundary modeling; (3) a multistage refinement mechanism that progressively corrects errors in low-quality industrial imagery. An inference pipeline further integrates generated prompt-based automatic segmentation with human-in-the-loop interactive mask refinement. Experiments conducted on four real-world CT and X-ray datasets demonstrate consistent and significant performance gains over classical CNN baselines, recent segmentation architectures, and multiple SAM-family variants. Ablation studies verify the complementary roles of multiscale fusion and multistage refinement in enhancing robustness. These results indicate an accurate, generalizable, and annotation-efficient segmentation solution for safety-critical overhang inspection under degraded radiographic images.